Digital Signal Processing 1292025735, 9781292025735

A significant revision of a best-selling text for the introductory digital signal processing course. This book presents

482 98 9MB

English Pages 1018 [1019] Year 2013

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Cover......Page 1
Table of Contents......Page 4
1. Introduction......Page 6
2. Discrete-Time Signals and Systems......Page 48
3. The z-Transform and Its Application to the Analysis of LTI Systems......Page 156
4. Frequency Analysis of Signals......Page 234
5. Frequency-Domain Analysis of LTI Systems......Page 312
6. Sampling and Reconstruction of Signals......Page 400
7. The Discrete Fourier Transform: Its Properties and Applications......Page 466
8. Efficient Computation of the DFT: Fast Fourier Transform Algorithms......Page 528
9. Implementation of Discrete-Time Systems......Page 582
10. Design of Digital Filters......Page 674
11. Multirate Digital Signal Processing......Page 772
12. Linear Prediction and Optimum Linear Filters......Page 846
13. Adaptive Filters......Page 904
14. Appendix: Random Number Generators......Page 986
15. Appendix: Tables of Transition Coefficients for the Design of Lnear-Phase FIR Filters......Page 992
16. References and Bibliography......Page 998
C......Page 1012
D......Page 1013
G......Page 1014
M......Page 1015
P......Page 1016
S......Page 1017
V......Page 1018
Z......Page 1019
Recommend Papers

Digital Signal Processing
 1292025735, 9781292025735

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Digital Signal Processing John G. Proakis Dimitris K. Manolakis Fourth Edition

Pearson Education Limited Edinburgh Gate Harlow Essex CM20 2JE England and Associated Companies throughout the world Visit us on the World Wide Web at: www.pearsoned.co.uk © Pearson Education Limited 2014 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without either the prior written permission of the publisher or a licence permitting restricted copying in the United Kingdom issued by the Copyright Licensing Agency Ltd, Saffron House, 6–10 Kirby Street, London EC1N 8TS. All trademarks used herein are the property of their respective owners. The use of any trademark in this text does not vest in the author or publisher any trademark ownership rights in such trademarks, nor does the use of such trademarks imply any affiliation with or endorsement of this book by such owners.

ISBN 10: 1-292-02573-5 ISBN 13: 978-1-292-02573-5

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library Printed in the United States of America

P

E

A

R

S

O

N

C U

S T O

M

L

I

B

R

A

R Y

Table of Contents

1. Introduction John G. Proakis/Dimitris G. Manolakis

1

2. Discrete-Time Signals and Systems John G. Proakis/Dimitris G. Manolakis

43

3. The z-Transform and Its Application to the Analysis of LTI Systems John G. Proakis/Dimitris G. Manolakis

151

4. Frequency Analysis of Signals John G. Proakis/Dimitris G. Manolakis

229

5. Frequency-Domain Analysis of LTI Systems John G. Proakis/Dimitris G. Manolakis

307

6. Sampling and Reconstruction of Signals John G. Proakis/Dimitris G. Manolakis

395

7. The Discrete Fourier Transform: Its Properties and Applications John G. Proakis/Dimitris G. Manolakis

461

8. Efficient Computation of the DFT: Fast Fourier Transform Algorithms John G. Proakis/Dimitris G. Manolakis

523

9. Implementation of Discrete-Time Systems John G. Proakis/Dimitris G. Manolakis

577

10. Design of Digital Filters John G. Proakis/Dimitris G. Manolakis

669

11. Multirate Digital Signal Processing John G. Proakis/Dimitris G. Manolakis

767

12. Linear Prediction and Optimum Linear Filters John G. Proakis/Dimitris G. Manolakis

841

13. Adaptive Filters John G. Proakis/Dimitris G. Manolakis

899

I

14. Appendix: Random Number Generators John G. Proakis/Dimitris G. Manolakis

981

15. Appendix: Tables of Transition Coefficients for the Design of Lnear-Phase FIR Filters John G. Proakis/Dimitris G. Manolakis

987

16. References and Bibliography John G. Proakis/Dimitris G. Manolakis Index

II

993 1007

Introduction

Digital signal processing is an area of science and engineering that has developed rapidly over the past 40 years. This rapid development is a result of the significant advances in digital computer technology and integrated-circuit fabrication. The digital computers and associated digital hardware of four decades ago were relatively large and expensive and, as a consequence, their use was limited to general-purpose non-real-time (off-line) scientific computations and business applications. The rapid developments in integrated-circuit technology, starting with medium-scale integration (MSI) and progressing to large-scale integration (LSI), and now, very-large-scale integration (VLSI) of electronic circuits has spurred the development of powerful, smaller, faster, and cheaper digital computers and special-purpose digital hardware. These inexpensive and relatively fast digital circuits have made it possible to construct highly sophisticated digital systems capable of performing complex digital signal processing functions and tasks, which are usually too difficult and/or too expensive to be performed by analog circuitry or analog signal processing systems. Hence many of the signal processing tasks that were conventionally performed by analog means are realized today by less expensive and often more reliable digital hardware. We do not wish to imply that digital signal processing is the proper solution for all signal processing problems. Indeed, for many signals with extremely wide bandwidths, real-time processing is a requirement. For such signals, analog or, perhaps, optical signal processing is the only possible solution. However, where digital circuits are available and have sufficient speed to perform the signal processing, they are usually preferable. Not only do digital circuits yield cheaper and more reliable systems for signal processing, they have other advantages as well. In particular, digital processing hardware allows programmable operations. Through software, one can more eas-

From Chapter 1 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

1

Introduction

ily modify the signal processing functions to be performed by the hardware. Thus digital hardware and associated software provide a greater degree of flexibility in system design. Also, there is often a higher order of precision achievable with digital hardware and software compared with analog circuits and analog signal processing systems. For all these reasons, there has been an explosive growth in digital signal processing theory and applications over the past three decades. We begin by introducing some of with the process of converting an analog signal to digital form suitable for digital processing. As we shall see, digital processing of analog signals has some drawbacks. First, and foremost, conversion of an analog signal to digital form, accomplished by sampling the signal and quantizing the samples, results in a distortion that prevents us from reconstructing the original analog signal from the quantized samples. Control of the amount of this distortion is achieved by proper choice of the sampling rate and the precision in the quantization process. Second, there are finite precision effectsthat must be considered in the digital processing of the quantized samples.

1

Signals, Systems, and Signal Processing A signal is defined as any physical quantity that varies with time, space, or any other independent variable or variables. Mathematically, we describe a signal as a function of one or more independent variables. For example, the functions s1 (t) = 5t s2 (t) = 20t 2

(1.1)

describe two signals, one that varies linearly with the independent variable t (time) and a second that varies quadratically with t . As another example, consider the function s(x, y) = 3x + 2xy + 10y 2

(1.2)

This function describes a signal of two independent variables x and y that could represent the two spatial coordinates in a plane. The signals described by (1.1) and (1.2) belong to a class of signals that are precisely defined by specifying the functional dependence on the independent variable. However, there are cases where such a functional relationship is unknown or too highly complicated to be of any practical use. For example, a speech signal (see Fig. 1.1) cannot be described functionally by expressions such as (1.1). In general, a segment of speech may be represented to

2

Introduction

Figure 1.1

Example of a speech signal.

a high degree of accuracy as a sum of several sinusoids of different amplitudes and frequencies, that is, as N  Ai (t) sin[2π Fi (t)t + θi (t)] (1.3) i=1

where {Ai (t)}, {Fi (t)}, and {θi (t)} are the sets of (possibly time-varying) amplitudes, frequencies, and phases, respectively, of the sinusoids. In fact, one way to interpret the information content or message conveyed by any short time segment of the speech signal is to measure the amplitudes, frequencies, and phases contained in the short time segment of the signal. Another example of a natural signal is an electrocardiogram (ECG). Such a signal provides a doctor with information about the condition of the patient’s heart. Similarly, an electroencephalogram (EEG) signal provides information about the activity of the brain. Speech, electrocardiogram, and electroencephalogram signals are examples of information-bearing signals that evolve as functions of a single independent variable, namely, time. An example of a signal that is a function of two independent variables is an image signal. The independent variables in this case are the spatial coordinates. These are but a few examples of the countless number of natural signals encountered in practice. Associated with natural signals are the means by which such signals are generated. For example, speech signals are generated by forcing air through the vocal cords. Images are obtained by exposing a photographic film to a scene or an object. Thus signal generation is usually associated with a system that responds to a stimulus or force. In a speech signal, the system consists of the vocal cords and the vocal tract, also called the vocal cavity. The stimulus in combination with the system is called a signal source. Thus we have speech sources, images sources, and various other types of signal sources. A system may also be defined as a physical device that performs an operation on a signal. For example, a filter used to reduce the noise and interference corrupting a desired information-bearing signal is called a system. In this case the filter performs some operation(s) on the signal, which has the effect of reducing (filtering) the noise and interference from the desired information-bearing signal.

3

Introduction

When we pass a signal through a system, as in filtering, we say that we have processed the signal. In this case the processing of the signal involves filtering the noise and interference from the desired signal. In general, the system is characterized by the type of operation that it performs on the signal. For example, if the operation is linear, the system is called linear. If the operation on the signal is nonlinear, the system is said to be nonlinear, and so forth. Such operations are usually referred to as signal processing. For our purposes, it is convenient to broaden the definition of a system to include not only physical devices, but also software realizations of operations on a signal. In digital processing of signals on a digital computer, the operations performed on a signal consist of a number of mathematical operations as specified by a software program. In this case, the program represents an implementation of the system in software. Thus we have a system that is realized on a digital computer by means of a sequence of mathematical operations; that is, we have a digital signal processing system realized in software. For example, a digital computer can be programmed to perform digital filtering. Alternatively, the digital processing on the signal may be performed by digital hardware (logic circuits) configured to perform the desired specified operations. In such a realization, we have a physical device that performs the specified operations. In a broader sense, a digital system can be implemented as a combination of digital hardware and software, each of which performs its own set of specified operations. This text deals with the processing of signals by digital means, either in software or in hardware. Since many of the signals encountered in practice are analog, we must also consider the problem of converting an analog siganl into a digital signal for processing. Thus we deal primarily with digital systems. The operations performed by such a system can usually be specified mathematically. The method or set of rules for implementing the system by a program that performs the corresponding mathematical operations is called an algorithm. Usually, there are many ways or algorithms by which a system can be implemented, either in software or in hardware, to perform the desired operations and computations. In practice, we have an interest in devising algorithms that are computationally efficient, fast, and easily implemented. Thus a major topic in the study of digital signal processing is the discussion of efficient algorithms for performing such operations as filtering, correlation, and spectral analysis.

1.1

Basic Elements of a Digital Signal Processing System

Most of the signals encountered in science and engineering are analog in nature. That is, the signals are functions of a continuous variable, such as time or space, and usually take on values in a continuous range. Such signals may be processed directly by appropriate analog systems (such as filters, frequency analyzers, or frequency multipliers) for the purpose of changing their characteristics or extracting some desired information. In such a case we say that the signal has been processed directly in its analog form, as illustrated in Fig. 1.2. Both the input signal and the output signal are in analog form.

4

Introduction

Figure 1.2

Analog output signal

Analog signal processor

Analog input signal

Analog signal processing.

Digital signal processing provides an alternative method for processing the analog signal, as illustrated in Fig. 1.3. To perform the processing digitally, there is a need for an interface between the analog signal and the digital processor. This interface is called an analog-to-digital (A/D) converter. The output of the A/D converter is a digital signal that is appropriate as an input to the digital processor. The digital signal processor may be a large programmable digital computer or a small microprocessor programmed to perform the desired operations on the input signal. It may also be a hardwired digital processor configured to perform a specified set of operations on the input signal. Programmable machines provide the flexibility to change the signal processing operations through a change in the software, whereas hardwired machines are difficult to reconfigure. Consequently, programmable signal processors are in very common use. On the other hand, when signal processing operations are well defined, a hardwired implementation of the operations can be optimized, resulting in a cheaper signal processor and, usually, one that runs faster than its programmable counterpart. In applications where the digital output from the digital signal processor is to be given to the user in analog form, such as in speech communications, we must provide another interface from the digital domain to the analog domain. Such an interface is called a digital-to-analog (D/A) converter. Thus the signal is provided to the user in analog form, as illustrated in the block diagram of Fig. 1.3. However, there are other practical applications involving signal analysis, where the desired information is conveyed in digital form and no D/A converter is required. For example, in the digital processing of radar signals, the information extracted from the radar signal, such as the position of the aircraft and its speed, may simply be printed on paper. There is no need for a D/A converter in this case.

1.2

Advantages of Digital over Analog Signal Processing

There are many reasons why digital signal processing of an analog signal may be preferable to processing the signal directly in the analog domain, as mentioned briefly earlier. First, a digital programmable system allows flexibility in reconfiguring the digital signal processing operations simply by changing the program. Reconfigu-

Analog input signal

Digital signal processor

A/D converter

Digital input signal

D/A converter

Analog output signal

Digital output signal

Figure 1.3 Block diagram of a digital signal processing system.

5

Introduction

ration of an analog system usually implies a redesign of the hardware followed by testing and verification to see that it operates properly. Accuracy considerations also play an important role in determining the form of the signal processor. Tolerances in analog circuit components make it extremely difficult for the system designer to control the accuracy of an analog signal processing system. On the other hand, a digital system provides much better control of accuracy requirements. Such requirements, in turn, result in specifying the accuracy requirements in the A/D converter and the digital signal processor, in terms of word length, floating-point versus fixed-point arithmetic, and similar factors. Digital signals are easily stored on magnetic media (tape or disk) without deterioration or loss of signal fidelity beyond that introduced in the A/D conversion. As a consequence, the signals become transportable and can be processed off-line in a remote laboratory. The digital signal processing method also allows for the implementation of more sophisticated signal processing algorithms. It is usually very difficult to perform precise mathematical operations on signals in analog form but these same operations can be routinely implemented on a digital computer using software. In some cases a digital implementation of the signal processing system is cheaper than its analog counterpart. The lower cost may be due to the fact that the digital hardware is cheaper, or perhaps it is a result of the flexibility for modifications provided by the digital implementation. As a consequence of these advantages, digital signal processing has been applied in practical systems covering a broad range of disciplines. We cite, for example, the application of digital signal processing techniques in speech processing and signal transmission on telephone channels, in image processing and transmission, in seismology and geophysics, in oil exploration, in the detection of nuclear explosions, in the processing of signals received from outer space, and in a vast variety of other applications. As already indicated, however, digital implementation has its limitations. One practical limitation is the speed of operation of A/D converters and digital signal processors. We shall see that signals having extremely wide bandwidths require fast-sampling-rate A/D converters and fast digital signal processors. Hence there are analog signals with large bandwidths for which a digital processing approach is beyond the state of the art of digital hardware.

2

Classification of Signals The methods we use in processing a signal or in analyzing the response of a system to a signal depend heavily on the characteristic attributes of the specific signal. There are techniques that apply only to specific families of signals. Consequently, any investigation in signal processing should start with a classification of the signals involved in the specific application.

2.1

Multichannel and Multidimensional Signals

As explained in Section 1, a signal is described by a function of one or more independent variables. The value of the function (i.e., the dependent variable) can be

6

Introduction

a real-valued scalar quantity, a complex-valued quantity, or perhaps a vector. For example, the signal s1 (t) = A sin 3π t is a real-valued signal. However, the signal s2 (t) = Aej 3πt = A cos 3π t + j A sin 3π t is complex valued. In some applications, signals are generated by multiple sources or multiple sensors. Such signals, in turn, can be represented in vector form. Figure 2.1 shows the three components of a vector signal that represents the ground acceleration due to an earthquake. This acceleration is the result of three basic types of elastic waves. The primary (P) waves and the secondary (S) waves propagate within the body of

Figure 2.1 Three components of ground acceleration measured a few kilometers from

the epicenter of an earthquake. (From Earthquakes, by B. A. Bold, ©1988 by W. H. Freeman and Company. Reprinted with permission of the publisher.)

7

Introduction

y

I(x1, y1)

y1

Figure 2.2

Example of a two-dimensional signal.

0

x1

x

rock and are longitudinal and transversal, respectively. The third type of elastic wave is called the surface wave, because it propagates near the ground surface. If sk (t), k = 1, 2, 3, denotes the electrical signal from the kth sensor as a function of time, the set of p = 3 signals can be represented by a vector S3 (t), where   s1 (t) S3 (t) = s2 (t) s3 (t) We refer to such a vector of signals as a multichannel signal.In electrocardiography, for example, 3-lead and 12-lead electrocardiograms (ECG) are often used in practice, which result in 3-channel and 12-channel signals. Let us now turn our attention to the independent variable(s). If the signal is a function of a single independent variable, the signal is called a one-dimensional signal. On the other hand, a signal is called M -dimensional if its value is a function of M independent variables. The picture shown in Fig. 2.2 is an example of a two-dimensional signal, since the intensity or brightness I (x, y) at each point is a function of two independent variables. On the other hand, a black-and-white television picture may be represented as I (x, y, t) since the brightness is a function of time. Hence the TV picture may be treated as a three-dimensional signal. In contrast, a color TV picture may be described by three intensity functions of the form Ir (x, y, t), Ig (x, y, t), and Ib (x, y, t), corresponding to the brightness of the three principal colors (red, green, blue) as functions of time. Hence the color TV picture is a three-channel, three-dimensional signal, which can be represented by the vector   Ir (x, y, t) I(x, y, t) = Ig (x, y, t) Ib (x, y, t)

8

Introduction

In this text we deal mainly with single-channel, one-dimensional real- or complex-valued signals and we refer to them simply as signals. In mathematical terms these signals are described by a function of a single independent variable. Although the independent variable need not be time, it is common practice to use t as the independent variable. In many cases the signal processing operations and algorithms developed in this text for one-dimensional, single-channel signals can be extended to multichannel and multidimensional signals.

2.2

Continuous-Time Versus Discrete-Time Signals

Signals can be further classified into four different categories depending on the characteristics of the time (independent) variable and the values they take. Continuoustime signals or analog signals are defined for every value of time and they take on values in the continuous interval (a, b), where a can be −∞ and b can be ∞. Mathematically, these signals can be described by functions of a continuous variable. The speech waveform in Fig. 1.1 and the signals x1 (t) = cos π t , x2 (t) = e−|t| , −∞ < t < ∞ are examples of analog signals. Discrete-time signalsare defined only at certain specific values of time. These time instants need not be equidistant, but in practice they are usually taken at equally spaced intervals for computational convenience and mathematical tractability. The signal x(tn ) = e−|tn | , n = 0, ±1, ±2, . . . provides an example of a discrete-time signal. If we use the index n of the discretetime instants as the independent variable, the signal value becomes a function of an integer variable (i.e., a sequence of numbers). Thus a discrete-time signal can be represented mathematically by a sequence of real or complex numbers. To emphasize the discrete-time nature of a signal, we shall denote such a signal as x(n) instead of x(t). If the time instants tn are equally spaced (i.e., tn = nT ), the notation x(nT ) is also used. For example, the sequence  n 0.8 , if n ≥ 0 (2.1) x(n) = 0, otherwise is a discrete-time signal, which is represented graphically as in Fig. 2.3. In applications, discrete-time signals may arise in two ways: 1. By selecting values of an analog signal at discrete-time instants. This process is called sampling and is discussed in more detail in Section 4. All measuring instruments that take measurements at a regular interval of time provide x(n)

1

… −1

0

1

2

3

4

5

6

7



n

Figure 2.3 Graphical representation of the discrete time signal

x(n) = 0.8n for n > 0 and x(n) = 0 for n < 0.

9

Introduction

Number of sunspots

200

100

0 1770

1790

1810

1830

1850

1870

Year

Figure 2.4 Wölfer annual sunspot numbers (1770–1869).

discrete-time signals. For example, the signal x(n) in Fig. 2.3 can be obtained by sampling the analog signal x(t) = 0.8t , t ≥ 0 and x(t) = 0, t < 0 once every second. 2. By accumulating a variable over a period of time. For example, counting the number of cars using a given street every hour, or recording the value of gold every day, results in discrete-time signals. Figure 2.4 shows a graph of the Wölfer sunspot numbers. Each sample of this discrete-time signal provides the number of sunspots observed during an interval of 1 year.

2.3 Continuous-Valued Versus Discrete-Valued Signals The values of a continuous-time or discrete-time signal can be continuous or discrete. If a signal takes on all possible values on a finite or an infinite range, it is said to be a continuous-valued signal. Alternatively, if the signal takes on values from a finite set of possible values, it is said to be a discrete-valued signal. Usually, these values are equidistant and hence can be expressed as an integer multiple of the distance between two successive values. A discrete-time signal having a set of discrete values is called a digital signal. Figure 2.5 shows a digital signal that takes on one of four possible values. In order for a signal to be processed digitally, it must be discrete in time and its values must be discrete (i.e., it must be a digital signal). If the signal to be processed is in analog form, it is converted to a digital signal by sampling the analog signal at discrete instants in time, obtaining a discrete-time signal, and then by quantizing its values to a set of discrete values, as described later in the chapter. The process

10

Introduction

x(n)





−1

0

1

2

3

4

5

6

7

8

n

Figure 2.5 Digital signal with four different amplitude values.

of converting a continuous-valued signal into a discrete-valued signal, called quantization, is basically an approximation process. It may be accomplished simply by rounding or truncation. For example, if the allowable signal values in the digital signal are integers, say 0 through 15, the continuous-value signal is quantized into these integer values. Thus the signal value 8.58 will be approximated by the value 8 if the quantization process is performed by truncation or by 9 if the quantization process is performed by rounding to the nearest integer. An explanation of the analog-to-digital conversion process is given later in the chapter.

2.4

Deterministic Versus Random Signals

The mathematical analysis and processing of signals requires the availability of a mathematical description for the signal itself. This mathematical description, often referred to as the signal model, leads to another important classification of signals. Any signal that can be uniquely described by an explicit mathematical expression, a table of data, or a well-defined rule is called deterministic. This term is used to emphasize the fact that all past, present, and future values of the signal are known precisely, without any uncertainty. In many practical applications, however, there are signals that either cannot be described to any reasonable degree of accuracy by explicit mathematical formulas, or such a description is too complicated to be of any practical use. The lack of such a relationship implies that such signals evolve in time in an unpredictable manner. We refer to these signals as random.The output of a noise generator, the seismic signal of Fig. 2.1, and the speech signal in Fig. 1.1 are examples of random signals. The mathematical framework for the theoretical analysis of random signals is provided by the theory of probability and stochastic processes. It should be emphasized at this point that the classification of a real-world signal as deterministic or random is not always clear. Sometimes, both approaches lead to meaningful results that provide more insight into signal behavior. At other times, the wrong classification may lead to erroneous results, since some mathematical tools may apply only to deterministic signals while others may apply only to random signals. This will become clearer as we examine specific mathematical tools.

11

Introduction

3

The Concept of Frequency in Continuous-Time and Discrete-Time Signals The concept of frequency is familiar to students in engineering and the sciences. This concept is basic in, for example, the design of a radio receiver, a high-fidelity system, or a spectral filter for color photography. From physics we know that frequency is closely related to a specific type of periodic motion called harmonic oscillation, which is described by sinusoidal functions. The concept of frequency is directly related to the concept of time. Actually, it has the dimension of inverse time. Thus we should expect that the nature of time (continuous or discrete) would affect the nature of the frequency accordingly.

3.1

Continuous-Time Sinusoidal Signals

A simple harmonic oscillation is mathematically described by the following continuoustime sinusoidal signal: xa (t) = A cos(t + θ),

−∞ < t < ∞

(3.1)

shown in Fig. 3.1. The subscript a used with x(t) denotes an analog signal. This signal is completely characterized by three parameters: A is the amplitude of the sinusoid,  is the frequency in radians per second (rad/s), and θ is the phase in radians. Instead of , we often use the frequency F in cycles per second or hertz (Hz), where  = 2π F

(3.2)

In terms of F , (3.1) can be written as xa (t) = A cos(2π F t + θ),

−∞ < t < ∞

(3.3)

We will use both forms, (3.1) and (3.3), in representing sinusoidal signals. xa(t) = A cos(2Ft + θ)

Tp = 1/F A A cos θ 0

Figure 3.1

Example of an analog sinusoidal signal.

12

t

Introduction

The analog sinusoidal signal in (3.3) is characterized by the following properties: A1. For every fixed value of the frequency F , xa (t) is periodic. Indeed, it can easily be shown, using elementary trigonometry, that xa (t + Tp ) = xa (t) where Tp = 1/F is the fundamental period of the sinusoidal signal. A2. Continuous-time sinusoidal signals with distinct (different) frequencies are themselves distinct. A3. Increasing the frequency F results in an increase in the rate of oscillation of the signal, in the sense that more periods are included in a given time interval. We observe that for F = 0, the value Tp = ∞ is consistent with the fundamental relation F = 1/Tp . Due to continuity of the time variable t , we can increase the frequency F , without limit, with a corresponding increase in the rate of oscillation. The relationships we have described for sinusoidal signals carry over to the class of complex exponential signals xa (t) = Aej (t+θ)

(3.4)

This can easily be seen by expressing these signals in terms of sinusoids using the Euler identity e±j φ = cos φ ± j sin φ

(3.5)

By definition, frequency is an inherently positive physical quantity. This is obvious if we interpret frequency as the number of cycles per unit time in a periodic signal. However, in many cases, only for mathematical convenience, we need to introduce negative frequencies. To see this we recall that the sinusoidal signal (3.1) may be expressed as xa (t) = A cos(t + θ) =

A j (t+θ) A −j (t+θ) + e e 2 2

(3.6)

which follows from (3.5). Note that a sinusoidal signal can be obtained by adding two equal-amplitude complex-conjugate exponential signals, sometimes called phasors, illustrated in Fig. 3.2. As time progresses the phasors rotate in opposite directions with angular frequencies ± radians per second. Since a positive frequency corresponds to counterclockwise uniform angular motion, a negative frequency simply corresponds to clockwise angular motion. For mathematical convenience, we use both negative and positive frequencies throughout this text. Hence the frequency range for analog sinusoids is −∞ < F < ∞.

13

Introduction

Im Ω

A/2 Ωt + θ Re

Ωt + θ

Figure 3.2

A/2

Representation of a cosine function by a pair of complex-conjugate exponentials (phasors).

3.2



Discrete-Time Sinusoidal Signals

A discrete-time sinusoidal signal may be expressed as x(n) = A cos(ωn + θ),

−∞ < n < ∞

(3.7)

where n is an integer variable, called the sample number, A is the amplitude of the sinusoid, ω is the frequency in radians per sample, and θ is the phase in radians. If instead of ω we use the frequency variable f defined by ω ≡ 2πf

(3.8)

the relation (3.7) becomes x(n) = A cos(2πf n + θ),

−∞ < n < ∞

(3.9)

The frequency f has dimensions of cycles per sample. In Section 4, where we consider the sampling of analog sinusoids, we relate the frequency variable f of a discrete-time sinusoid to the frequency F in cycles per second for the analog sinusoid. For the moment we consider the discrete-time sinusoid in (3.7) independently of the continuous-time sinusoid given in (3.1). Figure 3.3 shows a sinusoid with 1 cycles per sample) and phase frequency ω = π/6 radians per sample (f = 12 θ = π/3. x(n) = A cos(ωn + θ) A



Figure 3.3

Example of a discrete-time sinusoidal signal (ω = π/6 and θ = π/3).

14

… 0

n

−A

Introduction

In contrast to continuous-time sinusoids, the discrete-time sinusoids are characterized by the following properties: B1. A discrete-time sinusoid is periodic only if its frequency f is a rational number. By definition, a discrete-time signal x(n) is periodic with period N (N > 0) if and only if x(n + N ) = x(n) for all n (3.10) The smallest value of N for which (3.10) is true is called the fundamental period. The proof of the periodicity property is simple. For a sinusoid with frequency f0 to be periodic, we should have cos[2πf0 (N + n) + θ] = cos(2πf0 n + θ) This relation is true if and only if there exists an integer k such that 2πf0 N = 2kπ or, equivalently, f0 =

k N

(3.11)

According to (3.11), a discrete-time sinusoidal signal is periodic only if its frequency f0 can be expressed as the ratio of two integers (i.e., f0 is rational). To determine the fundamental period N of a periodic sinusoid, we express its frequency f0 as in (3.11) and cancel common factors so that k and N are relatively prime. Then the fundamental period of the sinusoid is equal to N . Observe that a small change in frequency can result in a large change in the period. For example, note that f1 = 31/60 implies that N1 = 60, whereas f2 = 30/60 results in N2 = 2. B2. Discrete-time sinusoids whose frequencies are separated by an integer multiple of 2π are identical. To prove this assertion, let us consider the sinusoid cos(ω0 n + θ). It easily follows that cos[(ω0 + 2π)n + θ] = cos(ω0 n + 2π n + θ) = cos(ω0 n + θ)

(3.12)

As a result, all sinusoidal sequences xk (n) = A cos(ωk n + θ),

k = 0, 1, 2, . . .

(3.13)

where ωk = ω0 + 2kπ,

−π ≤ ω0 ≤ π

are indistinguishable (i.e., identical). Any sequence resulting from a sinusoid with a frequency |ω| > π , or |f | > 21 , is identical to a sequence obtained from a sinusoidal signal with frequency |ω| < π . Because of this similarity, we call the sinusoid having the frequency |ω| > π an alias of a corresponding sinusoid with frequency |ω| < π . Thus we regard frequencies in the range −π ≤ ω ≤ π , or − 21 ≤ f ≤ 21 , as unique

15

Introduction

and all frequencies |ω| > π , or |f | > 21 , as aliases. The reader should notice the difference between discrete-time sinusoids and continuous-time sinusoids, where the latter result in distinct signals for  or F in the entire range −∞ <  < ∞ or −∞ < F < ∞. B3. The highest rate of oscillation in a discrete-time sinusoid is attained when ω = π (or ω = −π ) or, equivalently, f = 21 (or f = − 21 ). To illustrate this property, let us investigate the characteristics of the sinusoidal signal sequence x(n) = cos ω0 n when the frequency varies from 0 to π . To simplify the argument, we take values 1 of ω0 = 0, π/8, π/4, π/2, π corresponding to f = 0, 16 , 18 , 41 , 21 , which result in periodic sequences having periods N = ∞, 16, 8, 4, 2, as depicted in Fig. 3.4. We note that the period of the sinusoid decreases as the frequency increases. In fact, we can see that the rate of oscillation increases as the frequency increases. To see what happens for π ≤ ω0 ≤ 2π , we consider the sinusoids with frequencies ω1 = ω0 and ω2 = 2π − ω0 . Note that as ω1 varies from π to 2π , ω2 varies from π to 0. It can be easily seen that x1 (n) = A cos ω1 n = A cos ω0 n x2 (n) = A cos ω2 n = A cos(2π − ω0 )n

(3.14)

= A cos(−ω0 n) = x1 (n) x(n)



ω0 = 0

… n

x(n) 1

ω0 =

π 8

x(n) ω0 =

π 4

1







… n

n

x(n) ω0 =

π 2

x(n) ω0 = π 1

1

n

Figure 3.4 Signal x(n) = cos ω 0 n for various values of the frequency ω 0.

16

n

Introduction

Hence ω2 is an alias of ω1 . If we had used a sine function instead of a cosine function, the result would basically be the same, except for a 180 ◦ phase difference between the sinusoids x1 (n) and x2 (n). In any case, as we increase the relative frequency ω0 of a discrete-time sinusoid from π to 2π , its rate of oscillation decreases. For ω0 = 2π the result is a constant signal, as in the case for ω0 = 0. Obviously, for ω0 = π (or f = 21 ) we have the highest rate of oscillation. As for the case of continuous-time signals, negative frequencies can be introduced as well for discrete-time signals. For this purpose we use the identity x(n) = A cos(ωn + θ) =

A j (ωn+θ) A −j (ωn+θ) e + e 2 2

(3.15)

Since discrete-time sinusoidal signals with frequencies that are separated by an integer multiple of 2π are identical, it follows that the frequencies in any interval ω1 ≤ ω ≤ ω1 + 2π constitute all the existing discrete-time sinusoids or complex exponentials. Hence the frequency range for discrete-time sinusoids is finite with duration 2π . Usually, we choose the range 0 ≤ ω ≤ 2π or −π ≤ ω ≤ π (0 ≤ f ≤ 1, − 21 ≤ f ≤ 21 ), which we call the fundamental range.

3.3 Harmonically Related Complex Exponentials Sinusoidal signals and complex exponentials play a major role in the analysis of signals and systems. In some cases we deal with sets of harmonically related complex exponentials (or sinusoids). These are sets of periodic complex exponentials with fundamental frequencies that are multiples of a single positive frequency. Although we confine our discussion to complex exponentials, the same properties clearly hold for sinusoidal signals. We consider harmonically related complex exponentials in both continuous time and discrete time. Continuous-time exponentials.

The basic signals for continuous-time, harmoni-

cally related exponentials are sk (t) = ej k0 t = ej 2πkF0 t

k = 0, ±1, ±2, . . .

(3.16)

We note that for each value of k, sk (t) is periodic with fundamental period 1/(kF0 ) = Tp /k or fundamental frequency kF0 . Since a signal that is periodic with period Tp /k is also periodic with period k(Tp /k) = Tp for any positive integer k, we see that all of the sk (t) have a common period of Tp . F0 is allowed to take any value and all members of the set are distinct, in the sense that if k1 = k2 , then sk1 (t) = sk2 (t). From the basic signals in (3.16) we can construct a linear combination of harmonically related complex exponentials of the form xa (t) =

∞  k=−∞

ck sk (t) =

∞ 

ck ej k0 t

(3.17)

k=−∞

where ck , k = 0, ±1, ±2, . . . are arbitrary complex constants. The signal xa (t) is periodic with fundamental period Tp = 1/F0 , and its representation in terms of

17

Introduction

(3.17) is called the Fourier series expansion for xa (t). The complex-valued constants are the Fourier series coefficients and the signal sk (t) is called the kth harmonic of xa (t). Since a discrete-time complex exponential is periodic if its relative frequency is a rational number, we choose f0 = 1/N and we define the sets of harmonically related complex exponentials by

Discrete-time exponentials.

sk (n) = ej 2πkf0 n ,

k = 0, ±1, ±2, . . .

(3.18)

In contrast to the continuous-time case, we note that sk+N (n) = ej 2πn(k+N)/N = ej 2πn sk (n) = sk(n) This means that, consistent with (3.10), there are only N distinct periodic complex exponentials in the set described by (3.18). Furthermore, all members of the set have a common period of N samples. Clearly, we can choose any consecutive N complex exponentials, say from k = n0 to k = n0 + N − 1, to form a harmonically related set with fundamental frequency f0 = 1/N . Most often, for convenience, we choose the set that corresponds to n0 = 0, that is, the set sk (n) = ej 2πkn/N ,

k = 0, 1, 2, . . . , N − 1

(3.19)

As in the case of continuous-time signals, it is obvious that the linear combination x(n) =

N−1 

ck sk (n) =

k=0

N−1 

ck ej 2πkn/N

(3.20)

k=0

results in a periodic signal with fundamental period N . As we shall see later, this is the Fourier series representation for a periodic discrete-time sequence with Fourier coefficients {ck }. The sequence sk (n) is called the kth harmonic of x(n). EXAMPLE 3.1 Stored in the memory of a digital signal processor is one cycle of the sinusoidal signal  x(n) = sin

2πn +θ N



where θ = 2πq/N , where q and N are integers. (a) Determine how this table of values can be used to obtain values of harmonically related sinusoids having the same phase. (b) Determine how this table can be used to obtain sinusoids of the same frequency but different phase.

18

Introduction

Solution. (a) Let xk (n) denote the sinusoidal signal sequence   2πnk +θ xk (n) = sin N This is a sinusoid with frequency fk = k/N , which is harmonically related to x(n). But xk (n) may be expressed as   2π(kn) +θ xk (n) = sin N = x(kn) Thus we observe that xk (0) = x(0), xk (1) = x(k), xk (2) = x(2k), and so on. Hence the sinusoidal sequence xk (n) can be obtained from the table of values of x(n) by taking every k th value of x(n), beginning with x(0). In this manner we can generate the values of all harmonically related sinusoids with frequencies fk = k/N for k = 0, 1, . . . , N − 1. (b) We can control the phase θ of the sinusoid with frequency fk = k/N by taking the first value of the sequence from memory location q = θN/2π , where q is an integer. Thus the initial phase θ controls the starting location in the table and we wrap around the table each time the index (kn) exceeds N .

4

Analog-to-Digital and Digital-to-Analog Conversion Most signals of practical interest, such as speech, biological signals, seismic signals, radar signals, sonar signals, and various communications signals such as audio and video signals, are analog. To process analog signals by digital means, it is first necessary to convert them into digital form, that is, to convert them to a sequence of numbers having finite precision. This procedure is called analog-to-digital (A/D) conversion, and the corresponding devices are called A/D converters (ADCs). Conceptually, we view A/D conversion as a three-step process. This process is illustrated in Fig. 4.1. 1. Sampling. This is the conversion of a continuous-time signal into a discrete-time signal obtained by taking “samples” of the continuous-time signal at discretetime instants. Thus, if xa (t) is the input to the sampler, the output is xa (nT ) ≡ x(n), where T is called the sampling interval. A/D converter

xa(t)

Analog signal

Figure 4.1

Sampler

x(n)

Discrete-time signal

Quantizer

xq(n)

Coder

Quantized signal

01011…

Digital signal

Basic parts of an analog-to-digital (A/D) converter.

19

Introduction

2. Quantization. This is the conversion of a discrete-time continuous-valued signal into a discrete-time, discrete-valued (digital) signal. The value of each signal sample is represented by a value selected from a finite set of possible values. The difference between the unquantized sample x(n) and the quantized output xq (n) is called the quantization error. 3. Coding. In the coding process, each discrete value xq (n) is represented by a b-bit binary sequence. Although we model the A/D converter as a sampler followed by a quantizer and coder, in practice the A/D conversion is performed by a single device that takes xa (t) and produces a binary-coded number. The operations of sampling and quantization can be performed in either order but, in practice, sampling is always performed before quantization. In many cases of practical interest (e.g., speech processing) it is desirable to convert the processed digital signals into analog form. (Obviously, we cannot listen to the sequence of samples representing a speech signal or see the numbers corresponding to a TV signal.) The process of converting a digital signal into an analog signal is known as digital-to-analog (D/A) conversion. All D/A converters “connect the dots” in a digital signal by performing some kind of interpolation, whose accuracy depends on the quality of the D/A conversion process. Figure 4.2 illustrates a simple form of D/A conversion, called a zero-order hold or a staircase approximation. Other approximations are possible, such as linearly connecting a pair of successive samples (linear interpolation), fitting a quadratic through three successive samples (quadratic interpolation), and so on. Is there an optimum (ideal) interpolator? For signals having a limited frequency content (finite bandwidth), the sampling theorem introduced in the following section specifies the optimum form of interpolation. Sampling and quantization are treated in this section. In particular, we demonstrate that sampling does not result in a loss of information, nor does it introduce distortion in the signal if the signal bandwidth is finite. In principle, the analog signal can be reconstructed from the samples, provided that the sampling rate is sufficiently high to avoid the problem commonly called aliasing. On the other hand, quantization

Figure 4.2

Zero-order hold digital-to-analog (D/A) conversion.

20

Introduction

is a noninvertible or irreversible process that results in signal distortion. We shall show that the amount of distortion is dependent on the accuracy, as measured by the number of bits, in the A/D conversion process. The factors affecting the choice of the desired accuracy of the A/D converter are cost and sampling rate. In general, the cost increases with an increase in accuracy and/or sampling rate.

4.1 Sampling of Analog Signals There are many ways to sample an analog signal. We limit our discussion to periodic or uniform sampling, which is the type of sampling used most often in practice. This is described by the relation x(n) = xa (nT ),

−∞ < n < ∞

(4.1)

where x(n) is the discrete-time signal obtained by “taking samples” of the analog signal xa (t) every T seconds. This procedure is illustrated in Fig. 4.3. The time interval T between successive samples is called the sampling period or sample interval and its reciprocal 1/T = Fs is called the sampling rate (samples per second) or the sampling frequency (hertz). Periodic sampling establishes a relationship between the time variables t and n of continuous-time and discrete-time signals, respectively. Indeed, these variables are linearly related through the sampling period T or, equivalently, through the sampling rate Fs = 1/T , as n t = nT = (4.2) Fs As a consequence of (4.2), there exists a relationship between the frequency variable F (or ) for analog signals and the frequency variable f (or ω) for discretetime signals. To establish this relationship, consider an analog sinusoidal signal of the form xa (t) = A cos(2π F t + θ) (4.3)

Analog signal

xa(t)

x(n) = xa(nT) Fs = 1/T

Discrete-time signal

Sampler xa(t)

xa(t)

x(n)

x(n) = xa(nT)

0

t

0

n 1 2 3 4 5 6 7 8 9 T 2T … 5T … 9T … t = nT

Figure 4.3 Periodic sampling of an analog signal.

21

Introduction

which, when sampled periodically at a rate Fs = 1/T samples per second, yields xa (nT ) ≡ x(n) = A cos(2π F nT + θ)   2π nF = A cos +θ Fs

(4.4)

If we compare (4.4) with (3.9), we note that the frequency variables F and f are linearly related as F (4.5) f = Fs or, equivalently, as ω = T

(4.6)

The relation in (4.5) justifies the name relative or normalized frequency, which is sometimes used to describe the frequency variable f . As (4.5) implies, we can use f to determine the frequency F in hertz only if the sampling frequency Fs is known. We recall from Section 3.1 that the ranges of the frequency variables F or  for continuous-time sinusoids are −∞ < F < ∞ −∞ <  < ∞

(4.7)

However, the situation is different for discrete-time sinusoids. From Section 3.2 we recall that 1 1 − 2Fmax where Fmax is the largest frequency component in the analog signal. With the sampling rate selected in this manner, any frequency component, say |Fi | < Fmax , in the analog signal is mapped into a discrete-time sinusoid with a frequency −

Fi 1 1 ≤ fi = ≤ 2 Fs 2

(4.20)

or, equivalently, −π ≤ ωi = 2πfi ≤ π

(4.21)

Since, |f | = or |ω| = π is the highest (unique) frequency in a discrete-time signal, the choice of sampling rate according to (4.19) avoids the problem of aliasing. 1 2

27

Introduction

In other words, the condition Fs > 2Fmax ensures that all the sinusoidal components in the analog signal are mapped into corresponding discrete-time frequency components with frequencies in the fundamental interval. Thus all the frequency components of the analog signal are represented in sampled form without ambiguity, and hence the analog signal can be reconstructed without distortion from the sample values using an “appropriate” interpolation (digital-to-analog conversion) method. The “appropriate” or ideal interpolation formula is specified by the sampling theorem. Sampling Theorem. If the highest frequency contained in an analog signal xa (t) is Fmax = B and the signal is sampled at a rate Fs > 2Fmax ≡ 2B , then xa (t) can be exactly recovered from its sample values using the interpolation function

g(t) =

sin 2π Bt 2π Bt

(4.22)

Thus xa (t) may be expressed as 

∞ 

xa (t) =

xa

n=−∞

n Fs

   n g t− Fs

(4.23)

where xa (n/Fs ) = xa (nT ) ≡ x(n) are the samples of xa (t). When the sampling of xa (t) is performed at the minimum sampling rate Fs = 2B , the reconstruction formula in (4.23) becomes xa (t) =

∞ 

xa

n=−∞

n sin 2π B(t − n/2B) 2B 2π B(t − n/2B)

(4.24)

The sampling rate FN = 2B = 2Fmax is called the Nyquist rate. Figure 4.6 illustrates the ideal D/A conversion process using the interpolation function in (4.22). As can be observed from either (4.23) or (4.24), the reconstruction of xa (t) from the sequence x(n) is a complicated process, involving a weighted sum of the interpolation function g(t) and its time-shifted versions g(t − nT ) for−∞< n < ∞, where the weighting factors are the samples x(n). Because of the complexity and the infinite number of samples required in (4.23) or (4.24), these reconstruction formulas are primarily of theoretical interest. xa(t)

sample of xa(t)

Figure 4.6

Ideal D/A conversion (interpolation).

28

t (n − 2)T

(n − 1)T

nT

(n + 1)T

Introduction

EXAMPLE 4.3 Consider the analog signal xa (t) = 3 cos 50πt + 10 sin 300πt − cos 100πt What is the Nyquist rate for this signal? Solution.

The frequencies present in the signal above are F1 = 25 Hz,

F2 = 150 Hz,

F3 = 50 Hz

Thus Fmax = 150 Hz and according to (4.19), Fs > 2Fmax = 300 Hz The Nyquist rate is FN = 2Fmax . Hence FN = 300 Hz Discussion. It should be observed that the signal component 10 sin 300πt , sampled at the Nyquist rate FN = 300, results in the samples 10 sin πn, which are identically zero. In other words, we are sampling the analog sinusoid at its zero-crossing points, and hence we miss this signal component completely. This situation does not occur if the sinusoid is offset in phase by some amount θ . In such a case we have 10 sin(300πt + θ) sampled at the Nyquist rate FN = 300 samples per second, which yields the samples 10 sin(πn + θ) = 10(sin πn cos θ + cos πn sin θ) = 10 sin θ cos πn = (−1)n 10 sin θ Thus if θ = 0 or π , the samples of the sinusoid taken at the Nyquist rate are not all zero. However, we still cannot obtain the correct amplitude from the samples when the phase θ is unknown. A simple remedy that avoids this potentially troublesome situation is to sample the analog signal at a rate higher than the Nyquist rate.

EXAMPLE 4.4 Consider the analog signal xa (t) = 3 cos 2000πt + 5 sin 6000πt + 10 cos 12,000πt (a) What is the Nyquist rate for this signal? (b) Assume now that we sample this signal using a sampling rate Fs = 5000 samples/s. What is the discrete-time signal obtained after sampling? (c) What is the analog signal ya (t) that we can reconstruct from the samples if we use ideal interpolation?

29

Introduction

Solution. (a) The frequencies existing in the analog signal are F1 = 1 kHz,

F2 = 3 kHz,

F3 = 6 kHz

Thus Fmax = 6 kHz, and according to the sampling theorem, Fs > 2Fmax = 12 kHz The Nyquist rate is FN = 12 kHz (b) Since we have chosen Fs = 5 kHz, the folding frequency is Fs = 2.5 kHz 2 and this is the maximum frequency that can be represented uniquely by the sampled signal. By making use of (4.2) we obtain  x(n) = xa (nT ) = xa = 3 cos 2π = 3 cos 2π = 3 cos 2π

1 5

1 5

1 5

n Fs



n + 5 sin 2π

3 5

n + 10 cos 2π

6 5

n

n + 5 sin 2π 1 − 25 n + 10 cos 2π 1 + 15 n n + 5 sin 2π − 25 n + 10 cos 2π 15 n

Finally, we obtain x(n) = 13 cos 2π

1 5

n − 5 sin 2π

2 5

n

The same result can be obtained using Fig. 4.4. Indeed, since Fs = 5 kHz, the folding frequency is Fs /2 = 2.5 kHz. This is the maximum frequency that can be represented uniquely by the sampled signal. From (4.17) we have F0 = Fk − kFs . Thus F0 can be obtained by subtracting from Fk an integer multiple of Fs such that −Fs /2 ≤ F0 ≤ Fs /2. The frequency F1 is less than Fs /2 and thus it is not affected by aliasing. However, the other two frequencies are above the folding frequency and they will be changed by the aliasing effect. Indeed, F2 = F2 − Fs = −2 kHz F3 = F3 − Fs = 1 kHz From (4.5) it follows that f1 = 15 , f2 = − 25 , and f3 = 15 , which are in agreement with the result above. (c) Since the frequency components at only 1 kHz and 2 kHz are present in the sampled signal, the analog signal we can recover is ya(t) = 13 cos 2000πt − 5 sin 4000πt which is obviously different from the original signal xa (t). This distortion of the original analog signal was caused by the aliasing effect, due to the low sampling rate used.

30

Introduction

Although aliasing is a pitfall to be avoided, there are two useful practical applications based on the exploitation of the aliasing effect. These applications are the stroboscope and the sampling oscilloscope. Both instruments are designed to operate as aliasing devices in order to represent high frequencies as low frequencies. To elaborate, consider a signal with high-frequency components confined to a given frequency band B1 < F < B2 , where B2 − B1 ≡ B is defined as the bandwidth of the signal. We assume that B 0)

if and only if x(n + N ) = x(n) for all n

(1.20)

The smallest value of N for which (1.20) holds is called the (fundamental) period. If there is no value of N that satisfies (1.20), the signal is called nonperiodic or aperiodic. We have already observed that the sinusoidal signal of the form x(n) = A sin 2πf0 n

(1.21)

is periodic when f0 is a rational number, that is, if f0 can be expressed as f0 =

k N

(1.22)

where k and N are integers. The energy of a periodic signal x(n) over a single period, say, over the interval 0 ≤ n ≤ N − 1, is finite if x(n) takes on finite values over the period. However, the energy of the periodic signal for −∞ ≤ n ≤ ∞ is infinite. On the other hand, the average power of the periodic signal is finite and it is equal to the average power over a single period. Thus if x(n) is a periodic signal with fundamental period N and takes on finite values, its power is given by P =

N−1 1  |x(n)|2 N

(1.23)

n=0

Consequently, periodic signals are power signals. Symmetric (even) and antisymmetric (odd) signals. A real-valued signal x(n) is called symmetric (even) if x(−n) = x(n) (1.24)

On the other hand, a signal x(n) is called antisymmetric (odd) if x(−n) = −x(n)

(1.25)

We note that if x(n) is odd, then x(0) = 0. Examples of signals with even and odd symmetry are illustrated in Fig. 1.8.

50

Discrete-Time Signals and Systems

x (n)

−4 −3 −2 −1 0 1 2 3 4

n

(a) x (n)

−5 −4 −3 −2 −1

0 1 2 3 4 5

n

(b)

Figure 1.8 Example of even (a) and odd (b) signals.

We wish to illustrate that any arbitrary signal can be expressed as the sum of two signal components, one of which is even and the other odd. The even signal component is formed by adding x(n) to x(−n) and dividing by 2, that is, xe (n) =

1 [x(n) + x(−n)] 2

(1.26)

Clearly, xe (n) satisfies the symmetry condition (1.24). Similarly, we form an odd signal component xo (n) according to the relation xo (n) =

1 [x(n) − x(−n)] 2

(1.27)

Again, it is clear that xo (n) satisfies (1.25); hence it is indeed odd. Now, if we add the two signal components, defined by (1.26) and (1.27), we obtain x(n) , that is, x(n) = xe (n) + xo (n)

(1.28)

Thus any arbitrary signal can be expressed as in (1.28).

51

Discrete-Time Signals and Systems

1.3 Simple Manipulations of Discrete-Time Signals In this section we consider some simple modifications or manipulations involving the independent variable and the signal amplitude (dependent variable). Transformation of the independent variable (time). A signal x(n) may be shifted in time by replacing the independent variable n by n − k, where k is an integer. If k is a positive integer, the time shift results in a delay of the signal by k units of time. If k is a negative integer, the time shift results in an advance of the signal by |k| units in time. EXAMPLE 1.2 A signal x(n) is graphically illustrated in Fig. 1.9(a). Show a graphical representation of the signals x(n − 3) and x(n + 2). Solution. The signal x(n − 3) is obtained by delaying x(n) by three units in time. The result is illustrated in Fig. 1.9(b). On the other hand, the signal x(n + 2) is obtained by advancing x(n) by two units in time. The result is illustrated in Fig. 1.9(c). Note that delay corresponds to shifting a signal to the right, whereas advance implies shifting the signal to the left on the time axis.

Figure 1.9

Graphical representation of a signal, and its delayed and advanced versions.

52

Discrete-Time Signals and Systems

If the signal x(n) is stored on magnetic tape or on a disk or, perhaps, in the memory of a computer, it is a relatively simple operation to modify the base by introducing a delay or an advance. On the other hand, if the signal is not stored but is being generated by some physical phenomenon in real time, it is not possible to advance the signal in time, since such an operation involves signal samples that have not yet been generated. Whereas it is always possible to insert a delay into signal samples that have already been generated, it is physically impossible to view the future signal samples. Consequently, in real-time signal processing applications, the operation of advancing the time base of the signal is physically unrealizable. Another useful modification of the time base is to replace the independent variable n by −n. The result of this operation is a folding or a reflection of the signal about the time origin n = 0. EXAMPLE 1.3 Show the graphical representation of the signals x(−n) and x(−n+2), where x(n) is the signal illustrated in Fig. 1.10(a).

Figure 1.10

Graphical illustration of the folding and shifting operations.

53

Discrete-Time Signals and Systems

Solution. The new signal y(n) = x(−n) is shown in Fig. 1.10(b). Note that y(0) = x(0), y(1) = x(−1), y(2) = x(−2), and so on. Also, y(−1) = x(1), y(−2) = x(2), and so on. Therefore, y(n) is simply x(n) reflected or folded about the time origin n = 0. The signal y(n) = x(−n + 2) is simply x(−n) delayed by two units in time. The resulting signal is illustrated in Fig. 1.10(c). A simple way to verify that the result in Fig. 1.10(c) is correct is to compute samples, such as y(0) = x(2), y(1) = x(1), y(2) = x(0), y(−1) = x(3), and so on.

It is important to note that the operations of folding and time delaying (or advancing) a signal are not commutative. If we denote the time-delay operation by TD and the folding operation by FD, we can write TDk [x(n)] = x(n − k),

k>0

FD[x(n)] = x(−n)

(1.29)

Now TDk {FD[x(n)]} = TDk [x(−n)] = x(−n + k)

(1.30)

FD{TDk [x(n)]} = FD[x(n − k)] = x(−n − k)

(1.31)

whereas Note that because the signs of n and k in x(n − k) and x(−n + k) are different, the result is a shift of the signals x(n) and x(−n) to the right by k samples, corresponding to a time delay. A third modification of the independent variable involves replacing n by µn, where µ is an integer. We refer to this time-base modification as time scaling or down-sampling. EXAMPLE 1.4 Show the graphical representation of the signal y(n) = x(2n), where x(n) is the signal illustrated in Fig. 1.11(a). Solution. We note that the signal y(n) is obtained from x(n) by taking every other sample from x(n), starting with x(0). Thus y(0) = x(0), y(1) = x(2), y(2) = x(4), . . . and y(−1) = x(−2), y(−2) = x(−4), and so on. In other words, we have skipped the odd-numbered samples in x(n) and retained the even-numbered samples. The resulting signal is illustrated in Fig. 1.11(b).

If the signal x(n) was originally obtained by sampling an analog signal xa (t), then x(n) = xa (nT ), where T is the sampling interval. Now, y(n) = x(2n) = xa (2T n). Hence the time-scaling operation described in Example 1.4 is equivalent to changing the sampling rate from 1/T to 1/2T , that is, to decreasing the rate by a factor of 2. This is a down-sampling operation. Addition, multiplication, and scaling of sequences. Amplitude modifications include addition, multiplication, and scaling of discrete-time signals.

54

Discrete-Time Signals and Systems

x(n)

←7←6←5←4 ←3←2←1 0 1 2 3 4 5 6

n

(a) y(n) = x(2n)

←3 ←4

n

←2←1 0 1 2 3

(b)

Figure 1.11 Graphical illustration of down-sampling operation.

Amplitude scaling of a signal by a constant A is accomplished by multiplying the value of every signal sample by A. Consequently, we obtain y(n) = Ax(n),

−∞ < n < ∞

The sum of two signals x1 (n) and x2 (n) is a signal y(n), whose value at any instant is equal to the sum of the values of these two signals at that instant, that is, y(n) = x1 (n) + x2 (n),

−∞ < n < ∞

The product of two signals is similarly defined on a sample-to-sample basis as y(n) = x1 (n)x2 (n),

2

−∞ < n < ∞

Discrete-Time Systems In many applications of digital signal processing we wish to design a device or an algorithm that performs some prescribed operation on a discrete-time signal. Such a device or algorithm is called a discrete-time system. More specifically, a discrete-time system is a device or algorithm that operates on a discrete-time signal, called the input or excitation, according to some well-defined rule, to produce another discrete-time

55

Discrete-Time Signals and Systems



… x(n) Input signal or excitation

Discrete-time System

y(n) Output signal or response

Figure 2.1 Block diagram representation of a discrete-time system.

signal called the output or response of the system. In general, we view a system as an operation or a set of operations performed on the input signal x(n) to produce the output signal y(n). We say that the input signal x(n) is transformed by the system into a signal y(n), and express the general relationship between x(n) and y(n) as y(n) ≡ T [x(n)]

(2.1)

where the symbol T denotes the transformation (also called an operator) or processing performed by the system on x(n) to produce y(n). The mathematical relationship in (2.1) is depicted graphically in Fig. 2.1. There are various ways to describe the characteristics of the system and the operation it performs on x(n) to produce y(n). In this chapter we shall be concerned with the time-domain characterization of systems. We shall begin with an input–output description of the system. The input–output description focuses on the behavior at the terminals of the system and ignores the detailed internal construction or realization of the system.

2.1 Input–Output Description of Systems The input–output description of a discrete-time system consists of a mathematical expression or a rule, which explicitly defines the relation between the input and output signals (input–output relationship). The exact internal structure of the system is either unknown or ignored. Thus the only way to interact with the system is by using its input and output terminals (i.e., the system is assumed to be a “black box” to the user). To reflect this philosophy, we use the graphical representation depicted in Fig. 2.1, and the general input–output relationship in (2.1) or, alternatively, the notation T x(n) −→ y(n) (2.2) which simply means that y(n) is the response of the system T to the excitation x(n). The following examples illustrate several different systems. EXAMPLE 2.1 Determine the response of the following sytems to the input signal  |n|, −3 ≤ n ≤ 3 x(n) = 0, otherwise

56

Discrete-Time Signals and Systems

(a) y(n) = x(n) (identity system) (b) y(n) = x(n − 1) (unit delay system) (c) y(n) = x(n + 1) (unit advance system) (d) y(n) =

1 [x(n + 1) + x(n) + x(n − 1)] (moving average filter) 3

(e) y(n) = median{x(n + 1), x(n), x(n − 1)} (median filter) (f) y(n) =

n 

x(k) = x(n) + x(n − 1) + x(n − 2) + · · · (accumulator)

(2.3)

k=−∞

Solution.

First, we determine explicitly the sample values of the input signal x(n) = {. . . , 0, 3, 2, 1, 0, 1, 2, 3, 0, . . .} ↑

Next, we determine the output of each system using its input–output relationship. (a) In this case the output is exactly the same as the input signal. Such a system is known as the identity system. (b) This system simply delays the input by one sample. Thus its output is given by x(n) = {. . . , 0, 3, 2, 1, 0, 1, 2, 3, 0, . . .} ↑

(c) In this case the system “advances” the input one sample into the future. For example, the value of the output at time n = 0 is y(0) = x(1). The response of this system to the given input is x(n) = {. . . , 0, 3, 2, 1, 0, 1, 2, 3, 0, . . .} ↑

(d) The output of this system at any time is the mean value of the present, the immediate past, and the immediate future samples. For example, the output at time n = 0 is y(0) =

1 2 1 [x(−1) + x(0) + x(1)] = [1 + 0 + 1] = 3 3 3

Repeating this computation for every value of n, we obtain the output signal 2 5 5 y(n) = {. . . , 0, 1, , 2, 1, , 1, 2, , 1, 0, . . .} 3 3 3 ↑

(e) This system selects as its output at time n the median value of the three input samples x(n − 1), x(n), and x(n + 1). Thus the response of this system to the input signal x(n) is y(n) = {0, 2, 2, 1, 1, 1, 2, 2, 0, 0, 0, . . .} ↑

(f) This system is basically an accumulator that computes the running sum of all the past input values up to present time. The response of this system to the given input is y(n) = {. . . , 0, 3, 5, 6, 6, 7, 9, 12, 0, . . .} ↑

57

Discrete-Time Signals and Systems

We observe that for several of the systems considered in Example 2.1 the output at time n = n0 depends not only on the value of the input at n = n0 [i.e., x(n0 )], but also on the values of the input applied to the system before and after n = n0 . Consider, for instance, the accumulator in the example. We see that the output at time n = n0 depends not only on the input at time n = n0 , but also on x(n) at times n = n0 − 1, n0 − 2, and so on. By a simple algebraic manipulation the input–output relation of the accumulator can be written as y(n) =

n 

x(k) =

k=−∞

n−1 

x(k) + x(n)

k=−∞

(2.4)

= y(n − 1) + x(n) which justifies the term accumulator. Indeed, the system computes the current value of the output by adding (accumulating) the current value of the input to the previous output value. There are some interesting conclusions that can be drawn by taking a close look into this apparently simple system. Suppose that we are given the input signal x(n) for n ≥ n0 , and we wish to determine the output y(n) of this system for n ≥ n0 . For n = n0 , n0 + 1, . . . , (2.4) gives y(n0 ) = y(n0 − 1) + x(n0 ) y(n0 + 1) = y(n0 ) + x(n0 + 1) and so on. Note that we have a problem in computing y(n0 ), since it depends on y(n0 − 1). However, n0 −1

y(n0 − 1) =



x(k)

k=−∞

that is, y(n0 − 1) “summarizes” the effect on the system from all the inputs which had been applied to the system before time n0 . Thus the response of the system for n ≥ n0 to the input x(n) that is applied at time n0 is the combined result of this input and all inputs that had been applied previously to the system. Consequently, y(n), n ≥ n0 is not uniquely determined by the input x(n) for n ≥ n0 . The additional information required to determine y(n) for n ≥ n0 is the initial condition y(n0 − 1). This value summarizes the effect of all previous inputs to the system. Thus the initial condition y(n0 − 1) together with the input sequence x(n) for n ≥ n0 uniquely determine the output sequence y(n) for n ≥ n0 . If the accumulator had no excitation prior to n0 , the initial condition is y(n0 − 1) = 0. In such a case we say that the system is initially relaxed. Since y(n0 − 1) = 0, the output sequence y(n) depends only on the input sequence x(n) for n ≥ n0 . It is customary to assume that every system is relaxed at n = −∞. In this case, if an input x(n) is applied at n = −∞, the corresponding output y(n) is solely and uniquely determined by the given input.

58

Discrete-Time Signals and Systems

EXAMPLE 2.2 The accumulator described by (2.30) is excited by the sequence x(n) = nu(n). Determine its output under the condition that: (a) It is initially relaxed [i.e., y(−1) = 0]. (b) Initially, y(−1) = 1. Solution.

The output of the system is defined as

y(n) =

n 

x(k) =

k=−∞

−1 

x(k) +

k=−∞

= y(−1) +

n 

n 

x(k)

k=0

x(k)

k=0

= y(−1) +

n(n + 1) 2

(a) If the system is initially relaxed, y(−1) = 0 and hence y(n) =

n(n + 1) , 2

n≥0

(b) On the other hand, if the initial condition is y(−1) = 1, then y(n) = 1 +

2.2

n(n + 1) n2 + n + 2 = , 2 2

n≥0

Block Diagram Representation of Discrete-Time Systems

It is useful at this point to introduce a block diagram representation of discrete-time systems. For this purpose we need to define some basic building blocks that can be interconnected to form complex systems. An adder. Figure 2.2 illustrates a system (adder) that performs the addition of two signal sequences to form another (the sum) sequence, which we denote as y(n). Note that it is not necessary to store either one of the sequences in order to perform the addition. In other words, the addition operation is memoryless. x 1( n ) y( n ) = x 1 ( n ) + x 2 ( n ) +

Figure 2.2

Graphical representation of an adder.

x 2( n )

59

Discrete-Time Signals and Systems

A constant multiplier. This operation is depicted by Fig. 2.3, and simply represents applying a scale factor on the input x(n). Note that this operation is also memoryless. Figure 2.3

Graphical representation of a constant multiplier.

x(n)

a

y ( n ) = a x ( n)

Figure 2.4 illustrates the multiplication of two signal sequences to form another (the product) sequence, denoted in the figure as y(n). As in the preceding two cases, we can view the multiplication operation as memoryless.

A signal multiplier.

x 1( n )



y ( n ) = x 1 (n ) x 2 ( n )

Figure 2.4

Graphical representation of a signal multiplier.

x 2 ( n)

A unit delay element. The unit delay is a special system that simply delays the signal passing through it by one sample. Figure 2.5 illustrates such a system. If the input signal is x(n), the output is x(n − 1). In fact, the sample x(n − 1) is stored in memory at time n − 1 and it is recalled from memory at time n to form

y(n) = x(n − 1) Thus this basic building block requires memory. The use of the symbol z−1 to denote the unit of delay will become apparent when discussing the z-transform. Figure 2.5

Graphical representation of the unit delay element.

x(n )

z −1

y(n ) = x( n − 1 )

In contrast to the unit delay, a unit advance moves the input x(n) ahead by one sample in time to yield x(n + 1). Figure 2.6 illustrates this operation, with the operator z being used to denote the unit advance. We observe that any such advance is physically impossible in real time, since, in fact, it involves looking into the future of the signal. On the other hand, if we store the signal in the memory of the computer, we can recall any sample at any time. In such a non-realtime application, it is possible to advance the signal x(n) in time. A unit advance element.

Figure 2.6

Graphical representation of the unit advance element.

x(n ) z

y(n ) = x( n ⫹ 1 )

EXAMPLE 2.3 Using basic building blocks introduced above, sketch the block diagram representation of the discrete-time system described by the input–output relation y(n) =

1 1 1 y(n − 1) + x(n) + x(n − 1) 4 2 2

where x(n) is the input and y(n) is the output of the system.

60

(2.5)

Discrete-Time Signals and Systems

Black box 0.5

z −1 x(n)

+

y( n )

+ z −1

0.5 0.25

(a) Black box z −1 x(n)

0.5 +

y( n )

+ z −1 0.25 (b)

Figure 2.7 Block diagram realizations of the system y(n) = 0.25y(n − 1) +

0.5x(n) + 0.5x(n − 1).

Solution. According to (2.5), the output y(n) is obtained by multiplying the input x(n) by 0.5, multiplying the previous input x(n − 1) by 0.5, adding the two products, and then adding the previous output y(n − 1) multiplied by 41 . Figure 2.7(a) illustrates this block diagram realization of the system. A simple rearrangement of (2.5), namely, y(n) =

1 1 y(n − 1) + [x(n) + x(n − 1)] 4 2

(2.6)

leads to the block diagram realization shown in Fig. 2.7(b). Note that if we treat “the system” from the “viewpoint” of an input–output or an external description, we are not concerned about how the system is realized. On the other hand, if we adopt an internal description of the system, we know exactly how the system building blocks are configured. In terms of such a realization, we can see that a system is relaxed at time n = n0 if the outputs of all the delays existing in the system are zero at n = n0 (i.e., all memory is filled with zeros).

2.3

Classification of Discrete-Time Systems

In the analysis as well as in the design of systems, it is desirable to classify the systems according to the general properties that they satisfy. In fact, the mathematical techniques developed in this chapter and future study for analyzing and designing discrete-time systems depend heavily on the general characteristics of the systems that are being considered. For this reason it is necessary for us to develop a number of properties or categories that can be used to describe the general characteristics of systems. We stress the point that for a system to possess a given property, the property must hold for every possible input signal to the system. If a property holds for some

61

Discrete-Time Signals and Systems

input signals but not for others, the system does not possess that property. Thus a counterexample is sufficient to prove that a system does not possess a property. However, to prove that the system has some property, we must prove that this property holds for every possible input signal. Static versus dynamic systems. A discrete-time system is called static or memoryless if its output at any instant n depends at most on the input sample at the same time, but not on past or future samples of the input. In any other case, the system is said to be dynamic or to have memory. If the output of a system at time n is completely determined by the input samples in the interval from n − N to n(N ≥ 0), the system is said to have memory of duration N . If N = 0, the system is static. If 0 < N < ∞, the system is said to have finite memory, whereas if N = ∞, the system is said to have infinite memory. The systems described by the following input–output equations

y(n) = ax(n)

(2.7)

y(n) = nx(n) + bx 3 (n)

(2.8)

are both static or memoryless. Note that there is no need to store any of the past inputs or outputs in order to compute the present output. On the other hand, the systems described by the following input–output relations y(n) = x(n) + 3x(n − 1) y(n) =

n 

(2.9)

x(n − k)

(2.10)

x(n − k)

(2.11)

k=0

y(n) =

∞  k=0

are dynamic systems or systems with memory. The systems described by (2.9) and (2.10) have finite memory , whereas the system described by (2.11) has infinite memory. We observe that static or memoryless systems are described in general by input– output equations of the form y(n) = T [x(n), n]

(2.12)

and they do not include delay elements (memory). We can subdivide the general class of systems into the two broad categories, time-invariant systems and time-variant systems. A system is called time-invariant if its input–output characteristics do not change with time. To elaborate, suppose that we have a system T in a relaxed state

Time-invariant versus time-variant systems.

62

Discrete-Time Signals and Systems

which, when excited by an input signal x(n), produces an output signal y(n). Thus we write y(n) = T [x(n)] (2.13) Now suppose that the same input signal is delayed by k units of time to yield x(n−k), and again applied to the same system. If the characteristics of the system do not change with time, the output of the relaxed system will be y(n − k). That is, the output will be the same as the response to x(n), except that it will be delayed by the same k units in time that the input was delayed. This leads us to define a timeinvariant or shift-invariant system as follows. Definition. A relaxed system T is time invariant or shift invariant if and only if T

x(n) −→ y(n) implies that T

x(n − k) −→ y(n − k)

(2.14)

for every input signal x(n) and every time shift k. To determine if any given system is time invariant, we need to perform the test specified by the preceding definition. Basically, we excite the system with an arbitrary input sequence x(n), which produces an output denoted as y(n). Next we delay the input sequence by some amount k and recompute the output. In general, we can write the output as y(n, k) = T [x(n − k)] Now if this output y(n, k) = y(n − k), for all possible values of k, the system is time invariant. On the other hand, if the output y(n, k) = y(n − k), even for one value of k, the system is time variant. EXAMPLE 2.4 Determine if the systems shown in Fig. 2.8 are time invariant or time variant. Solution. (a) This system is described by the input–output equations y(n) = T [x(n)] = x(n) − x(n − 1)

(2.15)

Now if the input is delayed by k units in time and applied to the system, it is clear from the block diagram that the output will be y(n, k) = x(n − k) − x(n − k − 1)

(2.16)

On the other hand, from (2.14) we note that if we delay y(n) by k units in time, we obtain y(n − k) = x(n − k) − x(n − k − 1) (2.17) Since the right-hand sides of (2.16) and (2.17) are identical, it follows that y(n, k) = y(n − k). Therefore, the system is time invariant.

63

Discrete-Time Signals and Systems

y ( n ) = x( n ) − x( n − 1)

x(n) + −

“Differentiator”

z ←1 (a) y ( n ) = n x( n )

x( n ) ⫻

“Time” multiplier n

(b) x( n )

y (n ) = x(− n ) T

“Folder”

(c) y( n) = x ( n ) cos ω 0 n

x(n) +

Figure 2.8

Modulator

Examples of a time-invariant (a) and some time-variant systems (b)–(d).

cos ω 0 n (d)

(b) The input–output equation for this system is y(n) = T [x(n)] = nx(n)

(2.18)

The response of this system to x(n − k) is y(n, k) = nx(n − k)

(2.19)

Now if we delay y(n) in (2.18) by k units in time, we obtain y(n − k) = (n − k)x(n − k) (2.20) = nx(n − k) − kx(n − k) This system is time variant, since y(n, k) = y(n − k). (c) This system is described by the input–output relation y(n) = T [x(n)] = x(−n)

(2.21)

The response of this system to x(n − k) is y(n, k) = T [x(n − k)] = x(−n − k)

(2.22)

Now, if we delay the output y(n), as given by (2.21), by k units in time, the result will be y(n − k) = x(−n + k) Since y(n, k) = y(n − k), the system is time variant.

64

(2.23)

Discrete-Time Signals and Systems

(d) The input–output equation for this system is y(n) = x(n) cos ω0 n

(2.24)

The response of this system to x(n − k) is y(n, k) = x(n − k) cos ω0 n

(2.25)

If the expression in (2.24) is delayed by k units and the result is compared to (2.25), it is evident that the system is time variant.

The general class of systems can also be subdivided into linear systems and nonlinear systems. A linear system is one that satisfies the superposition principle. Simply stated, the principle of superposition requires that the response of the system to a weighted sum of signals be equal to the corresponding weighted sum of the responses (outputs) of the system to each of the individual input signals. Hence we have the following definition of linearity.

Linear versus nonlinear systems.

Definition. A system is linear if and only if T [a1 x1 (n) + a2 x2 (n)] = a1 T [x1 (n)] + a2 T [x2 (n)]

(2.26)

for any arbitrary input sequences x1 (n) and x2 (n), and any arbitrary constants a1 and a2 . Figure 2.9 gives a pictorial illustration of the superposition principle. The superposition principle embodied in the relation (2.26) can be separated into two parts. First, suppose that a2 = 0. Then (2.26) reduces to x 1 (n ) a1 T

+

y( n)

x 2 (n ) a2 x 1 (n )

T

a1 y′( n) +

x 2 (n )

T

a2

Graphical representation of the superposition principle. T is linear if and only if y(n) = y  (n).

Figure 2.9

65

Discrete-Time Signals and Systems

T [a1 x1 (n)] = a1 T [x1 (n)] = a1 y1 (n)

(2.27)

where y1 (n) = T [x1 (n)] The relation (2.27) demonstrates the multiplicative or scaling property of a linear system. That is, if the response of the system to the input x1 (n) is y1 (n), the response to a1 x1 (n) is simply a1 y1 (n). Thus any scaling of the input results in an identical scaling of the corresponding output. Second, suppose that a1 = a2 = 1 in (2.26). Then T [x1 (n) + x2 (n)] = T [x1 (n)] + T [x1 (n)] = y1 (n) + y2 (n)

(2.28)

This relation demonstrates the additivity property of a linear system. The additivity and multiplicative properties constitute the superposition principle as it applies to linear systems. The linearity condition embodied in (2.26) can be extended arbitrarily to any weighted linear combination of signals by induction. In general, we have x(n) =

M−1 

T

ak xk (n) −→ y(n) =

M−1 

ak yk (n)

(2.29)

k = 1, 2, . . . , M − 1

(2.30)

k=1

k=1

where yk (n) = T [xk (n)],

We observe from (2.27) that if a1 = 0, then y(n) = 0. In other words, a relaxed, linear system with zero input produces a zero output. If a system produces a nonzero output with a zero input, the system may be either nonrelaxed or nonlinear. If a relaxed system does not satisfy the superposition principle as given by the definition above, it is called nonlinear. EXAMPLE 2.5 Determine if the systems described by the following input–output equations are linear or nonlinear. (a) y(n) = nx(n) (b) y(n) = x(n2 ) (c) y(n) = x 2 (n) (d) y(n) = Ax(n) + B (e) y(n) = ex(n) Solution. (a) For two input sequences x1 (n) and x2 (n), the corresponding outputs are y1 (n) = nx1 (n) (2.31) y2 (n) = nx2 (n) A linear combination of the two input sequences results in the output y3 (n) = T [a1 x1 (n) + a2 x2 (n)] = n[a1 x1 (n) + a2 x2 (n)] (2.32) = a1 nx1 (n) + a2 nx2 (n)

66

Discrete-Time Signals and Systems

On the other hand, a linear combination of the two outputs in (2.31) results in the output a1 y1 (n) + a2 y2 (n) = a1 nx1 (n) + a2 nx2 (n)

(2.33)

Since the right-hand sides of (2.32) and (2.33) are identical, the system is linear. (b) As in part (a), we find the response of the system to two separate input signals x1 (n) and x2 (n). The result is y1 (n) = x1 (n2 ) (2.34) y2 (n) = x2 (n2 ) The output of the system to a linear combination of x1 (n) and x2 (n) is y3 (n) = T [a1 x1 (n) + a2 x2 (n)] = a1 x1 (n2 ) + a2 x2 (n2 )

(2.35)

Finally, a linear combination of the two outputs in (2.34) yields a1 y1 (n) + a2 y2 (n) = a1 x1 (n2 ) + a2 x2 (n2 )

(2.36)

By comparing (2.35) with (2.36), we conclude that the system is linear. (c) The output of the system is the square of the input. (Electronic devices that have such an input–output characteristic are called square-law devices.) From our previous discussion it is clear that such a system is memoryless. We now illustrate that this system is nonlinear. The responses of the system to two separate input signals are y1 (n) = x12 (n)

(2.37)

y2 (n) = x22 (n) The response of the system to a linear combination of these two input signals is y3 (n) = T [a1 x1 (n) + a2 x2 (n)] = [a1 x1 (n) + a2 x2 (n)]2

(2.38)

= a12 x12 (n) + 2a1 a2 x1 (n)x2 (n) + a22 x22 (n) On the other hand, if the system is linear, it will produce a linear combination of the two outputs in (2.37), namely, a1 y1 (n) + a2 y2 (n) = a1 x12 (n) + a2 x22 (n)

(2.39)

Since the actual output of the system, as given by (2.38), is not equal to (2.39), the system is nonlinear.

67

Discrete-Time Signals and Systems

(d) Assuming that the system is excited by x1 (n) and x2 (n) separately, we obtain the corresponding outputs y1 (n) = Ax1 (n) + B (2.40) y2 (n) = Ax2 (n) + B A linear combination of x1 (n) and x2 (n) produces the output y3 (n) = T [a1 x1 (n) + a2 x2 (n)] = A[a1 x1 (n) + a2 x2 (n)] + B

(2.41)

= Aa1 x1 (n) + a2 Ax2 (n) + B On the other hand, if the system were linear, its output to the linear combination of x1 (n) and x2 (n) would be a linear combination of y1 (n) and y2 (n), that is, a1 y1 (n) + a2 y2 (n) = a1 Ax1 (n) + a1 B + a2 Ax2 (n) + a2 B

(2.42)

Clearly, (2.41) and (2.42) are different and hence the system fails to satisfy the linearity test. The reason that this system fails to satisfy the linearity test is not that the system is nonlinear (in fact, the system is described by a linear equation) but the presence of the constant B . Consequently, the output depends on both the input excitation and on the parameter B = 0. Hence, for B = 0, the system is not relaxed. If we set B = 0, the system is now relaxed and the linearity test is satisfied. (e) Note that the system described by the input–output equation y(n) = ex(n)

(2.43)

is relaxed. If x(n) = 0, we find that y(n) = 1. This is an indication that the system is nonlinear. This, in fact, is the conclusion reached when the linearity test is applied.

Causal versus noncausal systems. We begin with the definition of causal discretetime systems. Definition. A system is said to be causal if the output of the system at any time n [i.e., y(n)] depends only on present and past inputs [i.e., x(n), x(n − 1), x(n − 2), . . .], but does not depend on future inputs [i.e., x(n + 1), x(n + 2), . . .]. In mathematical terms, the output of a causal system satisfies an equation of the form

y(n) = F [x(n), x(n − 1), x(n − 2), . . .]

(2.44)

where F [·] is some arbitrary function. If a system does not satisfy this definition, it is called noncausal. Such a system has an output that depends not only on present and past inputs but also on future inputs. It is apparent that in real-time signal processing applications we cannot observe future values of the signal, and hence a noncausal system is physically unrealizable (i.e., it cannot be implemented). On the other hand, if the signal is recorded so that the processing is done off-line (nonreal time), it is possible to implement a noncausal system, since all values of the signal are available at the time of processing. This is often the case in the processing of geophysical signals and images.

68

Discrete-Time Signals and Systems

EXAMPLE 2.6 Determine if the systems described by the following input–output equations are causal or noncausal.  (a) y(n) = x(n)−x(n−1) (b) y(n) = nk=−∞ x(k) (c) y(n) = ax(n) (d) y(n) = x(n)+3x(n+4) (e) y(n) = x(n2 ) (f) y(n) = x(2n) (g) y(n) = x(−n) Solution. The systems described in parts (a), (b), and (c) are clearly causal, since the output depends only on the present and past inputs. On the other hand, the systems in parts (d), (e), and (f) are clearly noncausal, since the output depends on future values of the input. The system in (g) is also noncausal, as we note by selecting, for example, n = −1, which yields y(−1) = x(1). Thus the output at n = −1 depends on the input at n = 1, which is two units of time into the future.

Stable versus unstable systems. Stability is an important property that must be considered in any practical application of a system. Unstable systems usually exhibit erratic and extreme behavior and cause overflow in any practical implementation. Here, we define mathematically what we mean by a stable system, and later, in Section 3.6, we explore the implications of this definition for linear, time-invariant systems. Definition. An arbitrary relaxed system is said to be bounded input–bounded output (BIBO) stable if and only if every bounded input produces a bounded output. The condition that the input sequence x(n) and the output sequence y(n) are bounded is translated mathematically to mean that there exist some finite numbers, say Mx and My , such that

|x(n)| ≤ Mx < ∞,

|y(n)| ≤ My < ∞

(2.45)

for all n. If, for some bounded input sequence x(n), the output is unbounded (infinite), the system is classified as unstable. EXAMPLE 2.7 Consider the nonlinear system described by the input–output equation y(n) = y 2 (n − 1) + x(n) As an input sequence we select the bounded signal x(n) = Cδ(n) where C is a constant. We also assume that y(−1) = 0. Then the output sequence is y(0) = C,

y(1) = C 2 ,

y(2) = C 4 ,

...,

y(n) = C 2n

Clearly, the output is unbounded when 1 < |C| < ∞. Therefore, the system is BIBO unstable, since a bounded input sequence has resulted in an unbounded output.

2.4

Interconnection of Discrete-Time Systems

Discrete-time systems can be interconnected to form larger systems. There are two basic ways in which systems can be interconnected: in cascade (series) or in parallel. These interconnections are illustrated in Fig. 2.10. Note that the two interconnected systems are different.

69

Discrete-Time Signals and Systems

x (n )

T1

y 1 ( n)

y (n )

T2

Tc (a)

T1

y 1 ( n)

x (n )

+ T2

y 3 ( n)

y 2 ( n)

Figure 2.10

Cascade (a) and parallel (b) interconnections of systems.

Tp (b)

In the cascade interconnection the output of the first system is y1 (n) = T1 [x(n)]

(2.46)

and the output of the second system is y(n) = T2 [y1 (n)] = T2 {T1 [x(n)]}

(2.47)

We observe that systems T1 and T2 can be combined or consolidated into a single overall system (2.48) Tc ≡ T2 T1 Consequently, we can express the output of the combined system as y(n) = Tc [x(n)] In general, the order in which the operations T1 and T2 are performed is important. That is, T2 T1 = T1 T2 for arbitrary systems. However, if the systems T1 and T2 are linear and time invariant, then (a) Tc is time invariant and (b) T2 T1 = T1 T2 , that is, the order in which the systems process the signal is not important. T2 T1 and T1 T2 yield identical output sequences. The proof of (a) follows. The proof of (b) is given in Section 3.4. To prove time invariance, suppose that T1 and T2 are time invariant; then T1

x(n − k) −→ y1 (n − k)

70

Discrete-Time Signals and Systems

and

T2

y1 (n − k) −→ y(n − k) Thus

Tc =T2 T1

x(n − k) −→ y(n − k) and therefore, Tc is time invariant. In the parallel interconnection, the output of the system T1 is y1 (n) and the output of the system T2 is y2 (n). Hence the output of the parallel interconnection is y3 (n) = y1 (n) + y2 (n) = T1 [x(n)] + T2 [x(n)] = (T1 + T2 )[x(n)] = Tp [x(n)] where Tp = T1 + T2 . In general, we can use parallel and cascade interconnection of systems to construct larger, more complex systems. Conversely, we can take a larger system and break it down into smaller subsystems for purposes of analysis and implementation. We shall use these notions later, in the design and implementation of digital filters.

3

Analysis of Discrete-Time Linear Time-Invariant Systems In Section 2 we classified systems in accordance with a number of characteristic properties or categories, namely: linearity, causality, stability, and time invariance. Having done so, we now turn our attention to the analysis of the important class of linear, time-invariant (LTI) systems. In particular, we shall demonstrate that such systems are characterized in the time domain simply by their response to a unit sample sequence. We shall also demonstrate that any arbitrary input signal can be decomposed and represented as a weighted sum of unit sample sequences. As a consequence of the linearity and time-invariance properties of the system, the response of the system to any arbitrary input signal can be expressed in terms of the unit sample response of the system. The general form of the expression that relates the unit sample response of the system and the arbitrary input signal to the output signal, called the convolution sum or the convolution formula, is also derived. Thus we are able to determine the output of any linear, time-invariant system to any arbitrary input signal.

3.1 Techniques for the Analysis of Linear Systems There are two basic methods for analyzing the behavior or response of a linear system to a given input signal. One method is based on the direct solution of the input–output equation for the system, which, in general, has the form y(n) = F [y(n − 1), y(n − 2), . . . , y(n − N ), x(n), x(n − 1), . . . , x(n − M)]

71

Discrete-Time Signals and Systems

where F [·] denotes some function of the quantities in brackets. Specifically, for an LTI system, we shall see later that the general form of the input–output relationship is y(n) = −

N 

ak y(n − k) +

M 

bk x(n − k)

(3.1)

k=0

k=1

where {ak } and {bk } are constant parameters that specify the system and are independent of x(n) and y(n). The input–output relationship in (3.1) is called a difference equation and represents one way to characterize the behavior of a discrete-time LTI system. The solution of (3.1) is the subject of Section 4. The second method for analyzing the behavior of a linear system to a given input signal is first to decompose or resolve the input signal into a sum of elementary signals. The elementary signals are selected so that the response of the system to each signal component is easily determined. Then, using the linearity property of the system, the responses of the system to the elementary signals are added to obtain the total response of the system to the given input signal. This second method is the one described in this section. To elaborate, suppose that the input signal x(n) is resolved into a weighted sum of elementary signal components {xk (n)} so that x(n) =



(3.2)

ck xk (n)

k

where the {ck } are the set of amplitudes (weighting coefficients) in the decomposition of the signal x(n). Now suppose that the response of the system to the elementary signal component xk (n) is yk (n). Thus, yk (n) ≡ T [xk (n)]

(3.3)

assuming that the system is relaxed and that the response to ck xk (n) is ck yk (n), as a consequence of the scaling property of the linear system. Finally, the total response to the input x(n) is y(n) = T [x(n)] = T

 

 ck xk (n)

k

=



ck T [xk (n)]

(3.4)

k

=



ck yk (n)

k

In (3.4) we used the additivity property of the linear system. Although to a large extent, the choice of the elementary signals appears to be arbitrary, our selection is heavily dependent on the class of input signals that we wish to consider. If we place no restriction on the characteristics of the input signals,

72

Discrete-Time Signals and Systems

their resolution into a weighted sum of unit sample (impulse) sequences proves to be mathematically convenient and completely general. On the other hand, if we restrict our attention to a subclass of input signals, there may be another set of elementary signals that is more convenient mathematically in the determination of the output. For example, if the input signal x(n) is periodic with period N, a mathematically convenient set of elementary signals is the set of exponentials xk (n) = ej ωk n ,

k = 0, 1, . . . , N − 1

(3.5)

where the frequencies {ωk } are harmonically related, that is,  ωk =

2π N

k,

k = 0, 1, . . . , N − 1

(3.6)

The frequency 2π/N is called the fundamental frequency, and all higher-frequency components are multiples of the fundamental frequency component. This subclass of input signals is considered in more detail later. For the resolution of the input signal into a weighted sum of unit sample sequences, we must first determine the response of the system to a unit sample sequence and then use the scaling and multiplicative properties of the linear system to determine the formula for the output given any arbitrary input. This development is described in detail as follows.

3.2 Resolution of a Discrete-Time Signal into Impulses Suppose we have an arbitrary signal x(n) that we wish to resolve into a sum of unit sample sequences. To utilize the notation established in the preceding section, we select the elementary signals xk (n) to be xk (n) = δ(n − k)

(3.7)

where k represents the delay of the unit sample sequence. To handle an arbitrary signal x(n) that may have nonzero values over an infinite duration, the set of unit impulses must also be infinite, to encompass the infinite number of delays. Now suppose that we multiply the two sequences x(n) and δ(n − k). Since δ(n − k) is zero everywhere except at n = k, where its value is unity, the result of this multiplication is another sequence that is zero everywhere except at n = k, where its value is x(k), as illustrated in Fig. 3.1. Thus x(n)δ(n − k) = x(k)δ(n − k)

(3.8)

is a sequence that is zero everywhere except at n = k, where its value is x(k). If we repeat the multiplication of x(n) with δ(n − m), where m is another delay (m = k), the result will be a sequence that is zero everywhere except at n = m, where its value is x(m). Hence (3.9) x(n)δ(n − m) = x(m)δ(n − m)

73

Discrete-Time Signals and Systems

Multiplication of a signal x(n) with a shifted unit sample sequence.

Figure 3.1

In other words, each multiplication of the signal x(n) by a unit impulse at some delay k , [i.e., δ(n − k)], in essence picks out the single value x(k) of the signal x(n) at the delay where the unit impulse is nonzero. Consequently, if we repeat this multiplication over all possible delays, −∞ < k < ∞, and sum all the product sequences, the result will be a sequence equal to the sequence x(n), that is, x(n) =

∞ 

x(k)δ(n − k)

(3.10)

k=−∞

We emphasize that the right-hand side of (3.10) is the summation of an infinite number of scaled unit sample sequences where the unit sample sequence δ(n−k) has an amplitude value of x(k). Thus the right-hand side of (3.10) gives the resolution or decomposition of any arbitrary signal x(n) into a weighted (scaled) sum of shifted unit sample sequences. EXAMPLE 3.1 Consider the special case of a finite-duration sequence given as x(n) = {2, 4, 0, 3} ↑

Resolve the sequence x(n) into a sum of weighted impulse sequences.

74

Discrete-Time Signals and Systems

Solution. Since the sequence x(n) is nonzero for the time instants n = −1, 0, 2, we need three impulses at delays k = −1, 0, Following (3.10) we find that x(n) = 2δ(n + 1) + 4δ(n) + 3δ(n − 2)

3.3

Response of LTI Systems to Arbitrary Inputs: The Convolution Sum

Having resolved an arbitrary input signal x(n) into a weighted sum of impulses, we are now ready to determine the response of any relaxed linear system to any input signal. First, we denote the response y(n, k) of the system to the input unit sample sequence at n = k by the special symbol h(n, k), −∞ < k < ∞. That is, y(n, k) ≡ h(n, k) = T [δ(n − k)]

(3.11)

In (3.11) we note that n is the time index and k is a parameter showing the location of the input impulse. If the impulse at the input is scaled by an amount ck ≡ x(k), the response of the system is the correspondingly scaled output, that is, ck h(n, k) = x(k)h(n, k)

(3.12)

Finally, if the input is the arbitrary signal x(n) that is expressed as a sum of weighted impulses, that is, ∞  x(k)δ(n − k) (3.13) x(n) = k=−∞

then the response of the system to x(n) is the corresponding sum of weighted outputs, that is,  ∞   y(n) = T [x(n)] = T x(k)δ(n − k) k=−∞

=

∞ 

x(k)T [δ(n − k)]

(3.14)

k=−∞

=

∞ 

x(k)h(n, k)

k=−∞

Clearly, (3.14) follows from the superposition property of linear systems, and is known as the superposition summation. We note that (3.14) is an expression for the response of a linear system to any arbitrary input sequence x(n). This expression is a function of both x(n) and the responses h(n, k) of the system to the unit impulses δ(n − k) for −∞ < k < ∞. In deriving (3.14) we used the linearity property of the system but not its timeinvariance property. Thus the expression in (3.14) applies to any relaxed linear (time-variant) system.

75

Discrete-Time Signals and Systems

If, in addition, the system is time invariant, the formula in (3.14) simplifies considerably. In fact, if the response of the LTI system to the unit sample sequence δ(n) is denoted as h(n), that is, h(n) ≡ T [δ(n)]

(3.15)

then by the time-invariance property, the response of the system to the delayed unit sample sequence δ(n − k) is h(n − k) = T [δ(n − k)]

(3.16)

Consequently, the formula in (3.14) reduces to y(n) =

∞ 

x(k)h(n − k)

(3.17)

k=−∞

Now we observe that the relaxed LTI system is completely characterized by a single function h(n), namely, its response to the unit sample sequence δ(n). In contrast, the general characterization of the output of a time-variant, linear system requires an infinite number of unit sample response functions, h(n, k), one for each possible delay. The formula in (3.17) that gives the response y(n) of the LTI system as a function of the input signal x(n) and the unit sample (impulse) response h(n) is called a convolution sum. We say that the input x(n) is convolved with the impulse response h(n) to yield the output y(n). We shall now explain the procedure for computing the response y(n), both mathematically and graphically, given the input x(n) and the impulse response h(n) of the system. Suppose that we wish to compute the output of the system at some time instant, say n = n0 . According to (3.17), the response at n = n0 is given as y(n0 ) =

∞ 

x(k)h(n0 − k)

(3.18)

k=−∞

Our first observation is that the index in the summation is k, and hence both the input signal x(k) and the impulse response h(n0 − k) are functions of k . Second, we observe that the sequences x(k) and h(n0 − k) are multiplied together to form a product sequence. The output y(n0 ) is simply the sum over all values of the product sequence. The sequence h(n0 − k) is obtained from h(k) by, first, folding h(k) about k = 0 (the time origin), which results in the sequence h(−k). The folded sequence is then shifted by n0 to yield h(n0 − k). To summarize, the process of computing the convolution between x(k) and h(k) involves the following four steps. 1. Folding. Fold h(k) about k = 0 to obtain h(−k). 2. Shifting. Shift h(−k) by n0 to the right (left) if n0 is positive (negative), to obtain h(n0 − k). 3. Multiplication. Multiply x(k) by h(n0 − k) to obtain the product sequence vn0 (k) ≡ x(k)h(n0 − k). 4. Summation. Sum all the values of the product sequence vn0 (k) to obtain the value of the output at time n = n0 .

76

Discrete-Time Signals and Systems

We note that this procedure results in the response of the system at a single time instant, say n = n0 . In general, we are interested in evaluating the response of the system over all time instants −∞ < n < ∞. Consequently, steps 2 through 4 in the summary must be repeated, for all possible time shifts −∞ < n < ∞. In order to gain a better understanding of the procedure for evaluating the convolution sum, we shall demonstrate the process graphically. The graphs will aid us in explaining the four steps involved in the computation of the convolution sum. EXAMPLE 3.2 The impulse response of a linear time-invariant system is h(n) = {1, 2, 1, −1} ↑

(3.19)

Determine the response of the system to the input signal x(n) = {1, 2, 3, 1} ↑

(3.20)

Solution. We shall compute the convolution according to the formula (3.17), but we shall use graphs of the sequences to aid us in the computation. In Fig. 3.2(a) we illustrate the input signal sequence x(k) and the impulse response h(k) of the system, using k as the time index in order to be consistent with (3.17). The first step in the computation of the convolution sum is to fold h(k). The folded sequence h(−k) is illustrated in Fig. 3.2(b). Now we can compute the output at n = 0, according to (3.17), which is ∞  x(k)h(−k) (3.21) y(0) = k=−∞

Since the shift n = 0, we use h(−k) directly without shifting it. The product sequence v0 (k) ≡ x(k)h(−k)

(3.22)

is also shown in Fig. 3.2(b). Finally, the sum of all the terms in the product sequence yields y(0) =

∞ 

h = −∞v0 (k) = 4

We continue the computation by evaluating the response of the system at n = 1. According to (3.17), ∞  y(1) = x(k)h(1 − k) (3.23) h=−∞

The sequence h(1 − k) is simply the folded sequence h(−k) shifted to the right by one unit in time. This sequence is illustrated in Fig. 3.2(c). The product sequence v1 (k) = x(k)h(1 − k) is also illustrated in Fig. 3.2(c). yields

(3.24)

Finally, the sum of all the values in the product sequence y(1) =

∞ 

v1 (k) = 8

k=−∞

77

Discrete-Time Signals and Systems

h(k)

x(k) 3

2 2 −1 0 1

k

3

−1 0 1 2 3 4

(a)

Fold

k

ly

ltip

Mu

ν0(k)

h(−k) 2

Product sequence

2

−2 −1 0 1 2

−1 0 1 2

k

k

(b) Shift

h(1 − k)

4

Multiply with x(k)

2

ν1(k)

Product sequence

−1 k

0 1 2

0 |1| 2

k

(c)

h(−1 − k) 2

−3

ν−1(k)

Multiply with x(k)

−2 −1 0 1

1

k

0 1 2 (d) 8 8 ∞ y(n) = Σ vn (k) k = −∞ 4 3

1

4 5

−3 −2 −1 0 1 2 3 −2

−1

6 7

(e)

Figure 3.2

78

Product sequence

Graphical computation of convolution.

n

k

Discrete-Time Signals and Systems

In a similar manner, we obtain y(2) by shifting h(−k) two units to the right, forming the product sequence v2 (k) = x(k)h(2−k) and then summing all the terms in the product sequence obtaining y(2) = 8. By shifting h(−k) farther to the right, multiplying the corresponding sequence, and summing over all the values of the resulting product sequences, we obtain y(3) = 3, y(4) = −2, y(5) = −1. For n > 5, we find that y(n) = 0 because the product sequences contain all zeros. Thus we have obtained the response y(n) for n > 0. Next we wish to evaluate y(n) for n < 0. We begin with n = −1. Then y(−1) =

∞ 

x(k)h(−1 − k)

(3.25)

k=−∞

Now the sequence h(−1 − k) is simply the folded sequence h(−k) shifted one time unit to the left. The resulting sequence is illustrated in Fig. 3.2(d). The corresponding product sequence is also shown in Fig. 3.2(d). Finally, summing over the values of the product sequence, we obtain y(−1) = 1 From observation of the graphs of Fig. 3.2, it is clear that any further shifts of h(−1 − k) to the left always result in an all-zero product sequence, and hence y(n) = 0

for n ≤ −2

Now we have the entire response of the system for −∞ < n < ∞, which we summarize below as (3.26) y(n) = {. . . , 0, 0, 1, 4, 8, 8, 3, −2, −1, 0, 0, . . .} ↑

In Example 3.2 we illustrated the computation of the convolution sum, using graphs of the sequences to aid us in visualizing the steps involved in the computation procedure. Before working out another example, we wish to show that the convolution operation is commutative in the sense that it is irrelevant which of the two sequences is folded and shifted. Indeed, if we begin with (3.17) and make a change in the variable of the summation, from k to m, by defining a new index m = n − k, then k = n − m and (3.17) becomes ∞ 

y(n) =

x(n − m)h(m)

(3.27)

m=−∞

Since m is a dummy index, we may simply replace m by k so that y(n) =

∞ 

x(n − k)h(k)

(3.28)

k=−∞

The expression in (3.28) involves leaving the impulse response h(k) unaltered, while the input sequence is folded and shifted. Although the output y(n) in (3.28)

79

Discrete-Time Signals and Systems

is identical to (3.17), the product sequences in the two forms of the convolution formula are not identical. In fact, if we define the two product sequences as vn (k) = x(k)h(n − k) wn (k) = x(n − k)h(k) it can be easily shown that vn (k) = wn (n − k) and therefore, y(n) =

∞ 

∞ 

vn (k) =

k=−∞

wn (n − k)

k=−∞

since both sequences contain the same sample values in a different arrangement. The reader is encouraged to rework Example 3.2 using the convolution sum in (3.28). EXAMPLE 3.3 Determine the output y(n) of a relaxed linear time-invariant system with impulse response h(n) = a nu (n), |a|1 when the input is a unit step sequence, that is, x(n) = u(n) Solution. In this case both h(n) and x(n) are infinite-duration sequences. We use the form of the convolution formula given by (3.28) in which x(k) is folded. The sequences h(k), x(k), and x(−k) are shown in Fig. 3.3. The product sequences v0 (k), v1 (k), and v2 (k) corresponding to x(−k) h(k), x(1 − k)h(k), and x(2 − k)h(k) are illustrated in Fig. 3.3(c), (d), and (e), respectively. Thus we obtain the outputs y(0) = 1 y(1) = 1 + a y(2) = 1 + a + a 2 Clearly, for n > 0, the output is y(n) = 1 + a + a 2 + · · · + a n =

(3.29)

1 − a n+1 1−a

On the other hand, for n < 0, the product sequences consist of all zeros. Hence y(n) = 0,

n n0 ]. Thus we obtain y(n0 ) =

∞ 

−1 

h(k)x(n0 − k) +

h(k)x(n0 − k)

k=−∞

k=0

= [h(0)x(n0 ) + h(1)x(n0 − 1) + h(2)x(n0 − 2) + · · ·] + [h(−1)x(n0 + 1) + h(−2)x(n0 + 2) + · · ·] We observe that the terms in the first sum involve x(n0 ), x(n0 − 1), . . . , which are the present and past values of the input signal. On the other hand, the terms in the second sum involve the input signal components x(n0 + 1), x(n0 + 2), . . . . Now, if the output at time n = n0 is to depend only on the present and past inputs, then, clearly, the impulse response of the system must satisfy the condition h(n) = 0,

n 0, a noncausal sequence. This terminology means that such a sequence could be the unit sample response of a causal or a noncausal system, respectively.

86

Discrete-Time Signals and Systems

If the input to a causal linear time-invariant system is a causal sequence [i.e., if x(n) = 0 for n < 0], the limits on the convolution formula are further restricted. In this case the two equivalent forms of the convolution formula become y(n) =

n 

h(k)x(n − k)

(3.41)

x(k)h(n − k)

(3.42)

k=0

=

n  k=0

We observe that in this case, the limits on the summations for the two alternative forms are identical, and the upper limit is growing with time. Clearly, the response of a causal system to a causal input sequence is causal, since y(n) = 0 for n < 0. EXAMPLE 3.5 Determine the unit step response of the linear time-invariant system with impulse response |a| < 1

h(n) = a n u(n),

Solution. Since the input signal is a unit step, which is a causal signal, and the system is also causal, we can use one of the special forms of the convolution formula, either (3.41) or (3.42). Since x(n) = 1 for n ≥ 0, (3.41) is simpler to use. Because of the simplicity of this problem, one can skip the steps involved with sketching the folded and shifted sequences. Instead, we use direct substitution of the signals sequences in (3.41) and obtain y(n) =

n 

ak

k=0

=

1 − a n+1 1−a

and y(n) = 0 for n < 0. We note that this result is identical to that obtained in Example 3.3. In this simple case, however, we computed the convolution algebraically without resorting to the detailed procedure outlined previously.

3.6 Stability of Linear Time-Invariant Systems As indicated previously, stability is an important property that must be considered in any practical implementation of a system. We defined an arbitrary relaxed system as BIBO stable if and only if its output sequence y(n) is bounded for every bounded input x(n). If x(n) is bounded, there exists a constant Mx such that |x(n)| ≤ Mx < ∞ Similarly, if the output is bounded, there exists a constant My such that |y(n)| < My < ∞ for all n.

87

Discrete-Time Signals and Systems

Now, given such a bounded input sequence x(n) to a linear time-invariant system, let us investigate the implications of the definition of stability on the characteristics of the system. Toward this end, we work again with the convolution formula y(n) =

∞ 

h(k)x(n − k)

k=−∞

If we take the absolute value of both sides of this equation, we obtain  ∞       |y(n)| =  h(k)x(n − k)   k=−∞

Now, the absolute value of the sum of terms is always less than or equal to the sum of the absolute values of the terms. Hence |y(n)| ≤

∞ 

|h(k)||x(n − k)|

k=−∞

If the input is bounded, there exists a finite number Mx such that |x(n)| ≤ Mx . By substituting this upper bound for x(n) in the equation above, we obtain |y(n)| ≤ Mx

∞ 

|h(k)|

k=−∞

From this expression we observe that the output is bounded if the impulse response of the system satisfies the condition Sh ≡

∞ 

|h(k)| < ∞

(3.43)

k=−∞

That is, a linear time-invariant system is stable if its impulse response is absolutely summable. This condition is not only sufficient but it is also necessary to ensure the stability of the system. Indeed, we shall show that if Sh = ∞, there is a bounded input for which the output is not bounded. We choose the bounded input  ∗ h (−n) , h(n) = 0 x(n) = |h(−n)| 0, h(n) = 0 where h∗ (n) is the complex conjugate of h(n). It is sufficient to show that there is one value of n for which y(n) is unbounded. For n = 0 we have y(0) =

∞  k=−∞

x(−k)h(k) =

∞  |h(k)|2 = Sh |h(k)|

k=−∞

Thus, if Sh = ∞, a bounded input produces an unbounded output since y(0) = ∞.

88

Discrete-Time Signals and Systems

The condition in (3.43) implies that the impulse response h(n) goes to zero as n approaches infinity. As a consequence, the output of the system goes to zero as n approaches infinity if the input is set to zero beyond n > n0 . To prove this, suppose that |x(n)| < Mx for n < n0 and x(n) = 0 for n ≥ n0 . Then, at n = n0 + N , the system output is y(n0 + N ) =

N−1 

h(k)x(n0 + N − k) +

k=−∞

∞ 

h(k)x(n0 + N − k)

k=N

But the first sum is zero since x(n) = 0 for n ≥ n0 . For the remaining part, we take the absolute value of the output, which is  ∞  ∞      |y(n0 + N )| =  h(k)x(n0 + N − k) ≤ |h(k)||x(n0 + N − k)|   k=N

≤ Mx

k=N

∞ 

|h(k)|

k=N

Now, as N approaches infinity, ∞ 

lim

N→∞

|h(n)| = 0

k=N

and hence lim |y(n0 + N )| = 0

N→∞

This result implies that any excitation at the input to the system, which is of a finite duration, produces an output that is “transient” in nature; that is, its amplitude decays with time and dies out eventually, when the system is stable. EXAMPLE 3.6 Determine the range of values of the parameter a for which the linear time-invariant system with impulse response h(n) = a n u(n) is stable. Solution. First, we note that the system is causal. Consequently, the lower index on the summation in (3.43) begins with k = 0. Hence ∞  k=0

|a k | =

∞ 

|a|k = 1 + |a| + |a|2 + · · ·

k=0

Clearly, this geometric series converges to ∞  k=0

|a|k =

1 1 − |a|

provided that |a| < 1. Otherwise, it diverges. Therefore, the system is stable if |a| < 1. Otherwise, it is unstable. In effect, h(n) must decay exponentially toward zero as n approaches infinity for the system to be stable.

89

Discrete-Time Signals and Systems

EXAMPLE 3.7 Determine the range of values of a and b for which the linear time-invariant system with impulse response  n a , n≥0 h(n) = bn , n < 0 is stable. Solution.

This system is noncasual. The condition on stability given by (3.43) yields ∞ 

|h(n)| =

n=−∞

∞  n=0

|a|n +

−1 

|b|n

n=−∞

From Example 3.6 we have already determined that the first sum converges for |a| < 1. The second sum can be manipulated as follows: −1  n=−∞

|b|n =

 ∞  1 1 1 1 = + · · · 1 + + |b|n |b| |b| |b|2 n=1

= β(1 + β + β 2 + · · ·) =

β 1−β

where β = 1/|b| must be less than unity for the geometric series to converge. Consequently, the system is stable if both |a| < 1 and |b| > 1 are satisfied.

3.7

Systems with Finite-Duration and Infinite-Duration Impulse Response

Up to this point we have characterized a linear time-invariant system in terms of its impulse response h(n). It is also convenient, however, to subdivide the class of linear time-invariant systems into two types, those that have a finite-duration impulse response (FIR) and those that have an infinite-duration impulse response (IIR). Thus an FIR system has an impulse response that is zero outside of some finite time interval. Without loss of generality, we focus our attention on causal FIR systems, so that h(n) = 0, n < 0 and n ≥ M The convolution formula for such a system reduces to y(n) =

M−1 

h(k)x(n − k)

k=0

A useful interpretation of this expression is obtained by observing that the output at any time n is simply a weighted linear combination of the input signal samples x(n), x(n − 1), . . . , x(n − M + 1). In other words, the system simply weights, by the values of the impulse response h(k), k = 0, 1, . . . , M − 1, the most recent M signal samples

90

Discrete-Time Signals and Systems

and sums the resulting M products. In effect, the system acts as a window that views only the most recent M input signal samples in forming the output. It neglects or simply “forgets” all prior input samples [i.e., x(n − M), x(n − M − 1), . . .]. Thus we say that an FIR system has a finite memory of length-M samples. In contrast, an IIR linear time-invariant system has an infinite-duration impulse response. Its output, based on the convolution formula, is y(n) =

∞ 

h(k)x(n − k)

k=0

where causality has been assumed, although this assumption is not necessary. Now, the system output is a weighted [by the impulse response h(k)] linear combination of the input signal samples x(n), x(n − 1), x(n − 2), . . . . Since this weighted sum involves the present and all the past input samples, we say that the system has an infinite memory. We investigate the characteristics of FIR and IIR systems in more detail.

4 Discrete-Time Systems Described by Difference Equations Up to this point we have treated linear and time-invariant systems that are characterized by their unit sample response h(n). In turn, h(n) allows us to determine the output y(n) of the system for any given input sequence x(n) by means of the convolution summation, ∞  h(k)x(n − k) (4.1) y(n) = k=−∞

In general, then, we have shown that any linear time-invariant system is characterized by the input–output relationship in (4.1). Moreover, the convolution summation formula in (4.1) suggests a means for the realization of the system. In the case of FIR systems, such a realization involves additions, multiplications, and a finite number of memory locations. Consequently, an FIR system is readily implemented directly, as implied by the convolution summation. If the system is IIR, however, its practical implementation as implied by convolution is clearly impossible, since it requires an infinite number of memory locations, multiplications, and additions. A question that naturally arises, then, is whether or not it is possible to realize IIR systems other than in the form suggested by the convolution summation. Fortunately, the answer is yes, there is a practical and computationally efficient means for implementing a family of IIR systems, as will be demonstrated in this section. Within the general class of IIR systems, this family of discrete-time systems is more conveniently described by difference equations. This family or subclass of IIR systems is very useful in a variety of practical applications, including the implementation of digital filters, and the modeling of physical phenomena and physical systems.

91

Discrete-Time Signals and Systems

4.1

Recursive and Nonrecursive Discrete-Time Systems

As indicated above, the convolution summation formula expresses the output of the linear time-invariant system explicitly and only in terms of the input signal. However, this need not be the case, as is shown here. There are many systems where it is either necessary or desirable to express the output of the system not only in terms of the present and past values of the input, but also in terms of the already available past output values. The following problem illustrates this point. Suppose that we wish to compute the cumulative average of a signal x(n) in the interval 0 ≤ k ≤ n, defined as 1  x(k), n+1 n

y(n) =

n = 0, 1, . . .

(4.2)

k=0

As implied by (4.2), the computation of y(n) requires the storage of all the input samples x(k) for 0 ≤ k ≤ n. Since n is increasing, our memory requirements grow linearly with time. Our intuition suggests, however, that y(n) can be computed more efficiently by utilizing the previous output value y(n − 1). Indeed, by a simple algebraic rearrangement of (4.2), we obtain (n + 1)y(n) =

n−1 

x(k) + x(n)

k=0

= ny(n − 1) + x(n) and hence y(n) =

1 n y(n − 1) + x(n) n+1 n+1

(4.3)

Now, the cumulative average y(n) can be computed recursively by multiplying the previous output value y(n − 1) by n/(n + 1), multiplying the present input x(n) by 1/(n + 1), and adding the two products. Thus the computation of y(n) by means of (4.3) requires two multiplications, one addition, and one memory location, as illustrated in Fig. 4.1. This is an example of a recursive system. In general, a system whose output y(n) at time n depends on any number of past output values y(n − 1), y(n − 2), . . . is called a recursive system. x(n)



+

1 n+1

Figure 4.1

Realization of a recursive cumulative averaging system.

92

z−1

⫻ n

y(n)

Discrete-Time Signals and Systems

To determine the computation of the recursive system in (4.3) in more detail, suppose that we begin the process with n = 0 and proceed forward in time. Thus, according to (4.3), we obtain y(0) = x(0) y(1) =

1 1 y(0) + x(1) 2 2

2 1 y(1) + x(2) 3 3 and so on. If one grows fatigued with this computation and wishes to pass the problem to someone else at some time, say n = n0 , the only information that one needs to provide his or her successor is the past value y(n0 − 1) and the new input samples x(n), x(n + 1), . . . . Thus the successor begins with y(2) =

y(n0 ) =

n0 1 y(n0 − 1) + x(n0 ) n0 + 1 n0 + 1

and proceeds forward in time until some time, say n = n1 , when he or she becomes fatigued and passes the computational burden to someone else with the information on the value y(n1 − 1), and so on. The point we wish to make in this discussion is that if one wishes to compute the response (in this case, the cumulative average) of the system (4.3) to an input signal x(n) applied at n = n0 , we need the value y(n0 − 1) and the input samples x(n) for n ≥ n0 . The term y(n0 − 1) is called the initial condition for the system in (4.3) and contains all the information needed to determine the response of the system for n ≥ n0 to the input signal x(n), independent of what has occurred in the past. The following example illustrates the use of a (nonlinear) recursive system to compute the square root of a number. EXAMPLE 4.1

Square-Root Algorithm

Many computers and calculators compute the square root of a positive number A, using the iterative algorithm  1 A , n = 0, 1, . . . sn−1 + sn = 2 sn−1 √ of A. As the iteration converges we have sn ≈ sn−1 . where sn−1 is an initial guess (estimate) √ Then it easily follows that sn ≈ A. Consider now the recursive system

x(n) 1 y(n − 1) + (4.4) y(n) = 2 y(n − 1) which is realized as in Fig. 4.2. If we excite this system with a step √ of amplitude A [i.e., A, the response y(n) of x(n) = Au(n)] and use as an initial condition y(−1) an estimate of √ the system will tend toward A as n increases. Note that in contrast to the system (4.3), we do not need to specify exactly the initial condition. A rough estimate is sufficient for the proper performance of the system. For example, if we let A = 2 and y(−1) = 1, we obtain y(0) = 23 , y(1) = 1.4166667, y(2) = 1.4142157. Similarly, for √ y(−1) = 1.5, we have y(0) = 1.416667, y(1) = 1.4142157. Compare these values with the 2, which is approximately 1.4142136.

93

Discrete-Time Signals and Systems

x(n)



+

1 2

y(n)

1 y(n − 1)

Figure 4.2

Realization of the square-root system.

y(n − 1)

z−1

We have now introduced two simple recursive systems, where the output y(n) depends on the previous output value y(n − 1) and the current input x(n). Both systems are causal. In general, we can formulate more complex causal recursive systems, in which the output y(n) is a function of several past output values and present and past inputs. The system should have a finite number of delays or, equivalently, should require a finite number of storage locations to be practically implemented. Thus the output of a causal and practically realizable recursive system can be expressed in general as y(n) = F [y(n − 1), y(n − 2), . . . , y(n − N ), x(n), x(n − 1), . . . , x(n − M)]

(4.5)

where F [·] denotes some function of its arguments. This is a recursive equation specifying a procedure for computing the system output in terms of previous values of the output and present and past inputs. In contrast, if y(n) depends only on the present and past inputs, then y(n) = F [x(n), x(n − 1), . . . , x(n − M)]

(4.6)

Such a system is called nonrecursive. We hasten to add that the causal FIR systems described in Section 3.7 in terms of the convolution sum formula have the form of (2.4.6). Indeed, the convolution summation for a causal FIR system is y(n) =

M 

h(k)x(n − k)

k=0

= h(0)x(n) + h(1)x(n − 1) + · · · + h(M)x(n − M) = F [x(n), x(n − 1), . . . , x(n − M)] where the function F [·] is simply a linear weighted sum of present and past inputs and the impulse response values h(n), 0 ≤ n ≤ M , constitute the weighting coefficients. Consequently, the causal linear time-invariant FIR systems described by the convolution formula in Section 3.7 are nonrecursive. The basic differences between nonrecursive and recursive systems are illustrated in Fig. 4.3. A simple inspection of this figure reveals that the fundamental difference between these two systems is the feedback loop in the recursive system, which feeds back the output of the system into the input. This feedback loop contains a delay element. The presence of this delay is crucial for the realizability of the system, since the absence of this delay would force the system to compute y(n) in terms of y(n), which is not possible for discrete-time systems.

94

Discrete-Time Signals and Systems

x(n)

F [x(n), x(n − 1), … , x(n − M)]

y(n)

(a)

x(n)

F [y(n − 1) … , y(n − N), x(n), … , x(n − M)]

z−1

Figure 4.3

Basic form for a causal and realizable (a) nonrecursive and (b) recursive system.

y(n)

(b)

The presence of the feedback loop or, equivalently, the recursive nature of (4.5) creates another important difference between recursive and nonrecursive systems. For example, suppose that we wish to compute the output y(n0 ) of a system when it is excited by an input applied at time n = 0. If the system is recursive, to compute y(n0 ), we first need to compute all the previous values y(0), y(1), . . . , y(n0 −1). In contrast, if the system is nonrecursive, we can compute the output y(n0 ) immediately without having y(n0 −1), y(n0 −2), . . . . In conclusion, the output of a recursive system should be computed in order [i.e., y(0), y(1), y(2), . . .], whereas for a nonrecursive system, the output can be computed in any order [i.e., y(200), y(15), y(3), y(300), etc.]. This feature is desirable in some practical applications.

4.2

Linear Time-Invariant Systems Characterized by Constant-Coefficient Difference Equations

In Section 3 we treated linear time-invariant systems and characterized them in terms of their impulse responses. In this subsection we focus our attention on a family of linear time-invariant systems described by an input–output relation called a difference equation with constant coeffficients. Systems described by constantcoefficient linear difference equations are a subclass of the recursive and nonrecursive systems introduced in the preceding subsection. To bring out the important ideas, we begin by treating a simple recursive system described by a first-order difference equation. Suppose that we have a recursive system with an input–output equation y(n) = ay(n − 1) + x(n)

(4.7)

where a is a constant. Figure 4.4 shows a block diagram realization of the system. In comparing this system with the cumulative averaging system described by the input–output equation (4.3), we observe that the system in (4.7) has a constant coefficient (independent of time), whereas the system described in (4.3) has timevariant coefficients. As we will show, (4.7) is an input–output equation for a linear timeinvariant system, whereas (4.3) describes a linear time-variant system.

95

Discrete-Time Signals and Systems

x(n)

y(n)

+

Figure 4.4

z−1

Block diagram realization of a simple recursive system.

a

Now, suppose that we apply an input signal x(n) to the system for n ≥ 0. We make no assumptions about the input signal for n < 0, but we do assume the existence of the initial condition y(−1). Since (4.7) describes the system output implicitly, we must solve this equation to obtain an explicit expression for the system output. Suppose that we compute successive values of y(n) for n ≥ 0, beginning with y(0). Thus y(0) = ay(−1) + x(0) y(1) = ay(0) + x(1) = a 2 y(−1) + ax(0) + x(1) y(2) = ay(1) + x(2) = a 3 y(−1) + a 2 x(0) + ax(1) + x(2) .. .

.. .

y(n) = ay(n − 1) + x(n) = a n+1 y(−1) + a n x(0) + a n−1 x(1) + · · · + ax(n − 1) + x(n) or, more compactly, y(n) = a n+1 y(−1) +

n 

a k x(n − k),

n≥0

(4.8)

k=0

The response y(n) of the system as given by the right-hand side of (4.8) consists of two parts. The first part, which contains the term y(−1), is a result of the initial condition y(−1) of the system. The second part is the response of the system to the input signal x(n). If the system is initially relaxed at time n = 0, then its memory (i.e., the output of the delay) should be zero. Hence y(−1) = 0. Thus a recursive system is relaxed if it starts with zero initial conditions. Because the memory of the system describes, in some sense, its “state,” we say that the system is at zero state and its corresponding output is called the zero-state response, and is denoted by yzs (n). Obviously, the zero-state response of the system (4.7) is given by yzs (n) =

n 

a k x(n − k),

n≥0

(4.9)

k=0

It is interesting to note that (4.9) is a convolution summation involving the input signal convolved with the impulse response h(n) = a n u(n)

96

(4.10)

Discrete-Time Signals and Systems

We also observe that the system described by the first-order difference equation in (4.7) is causal. As a result, the lower limit on the convolution summation in (4.9) is k = 0. Furthermore, the condition y(−1) = 0 implies that the input signal can be assumed causal and hence the upper limit on the convolution summation in (4.9) is n, since x(n − k) = 0 for k > n. In effect, we have obtained the result that the relaxed recursive system described by the first-order difference equation in (4.7) is a linear time-invariant IIR system with impulse response given by (4.10). Now, suppose that the system described by (4.7) is initially nonrelaxed [i.e., y(−1) = 0] and the input x(n) = 0 for all n. Then the output of the system with zero input is called the zero-input response or natural response and is denoted by yzi (n). From (4.7), with x(n) = 0 for −∞ < n < ∞, we obtain yzi (n) = a n+1 y(−1),

n≥0

(4.11)

We observe that a recursive system with nonzero initial condition is nonrelaxed in the sense that it can produce an output without being excited. Note that the zero-input response is due to the memory of the system. To summarize, the zero-input response is obtained by setting the input signal to zero, making it independent of the input. It depends only on the nature of the system and the initial condition. Thus the zero-input response is a characteristic of the system itself, and it is also known as the natural or free response of the system. On the other hand, the zero-state response depends on the nature of the system and the input signal. Since this output is a response forced upon it by the input signal, it is usually called the forced response of the system. In general, the total response of the system can be expressed as y(n) = yzi (n) + yzs (n). The system described by the first-order difference equation in (4.7) is the simplest possible recursive system in the general class of recursive systems described by linear constant-coefficient difference equations. The general form for such an equation is N M   ak y(n − k) + bk x(n − k) (4.12) y(n) = − k=0

k=1

or, equivalently, N  k=0

ak y(n − k) =

M 

bk x(n − k),

a0 ≡ 1

(4.13)

k=0

The integer N is called the order of the difference equation or the order of the system. The negative sign on the right-hand side of (4.12) is introduced as a matter of convenience to allow us to express the difference equation in (4.13) without any negative signs. Equation (4.12) expresses the output of the system at time n directly as a weighted sum of past outputs y(n − 1), y(n − 2), . . . , y(n − N ) as well as past and present input signals samples. We observe that in order to determine y(n) for n ≥ 0, we need the input x(n) for all n ≥ 0, and the initial conditions y(−1),

97

Discrete-Time Signals and Systems

y(−2), . . . , y(−N ). In other words, the initial conditions summarize all that we need to know about the past history of the response of the system to compute the present and future outputs. The general solution of the N-order constant-coefficient difference equation is considered in the following subsection. At this point we restate the properties of linearity, time invariance, and stability in the context of recursive systems described by linear constant-coefficient difference equations. As we have observed, a recursive system may be relaxed or nonrelaxed, depending on the initial conditions. Hence the definitions of these properties must take into account the presence of the initial conditions. We begin with the definition of linearity. A system is linear if it satisfies the following three requirements: 1. The total response is equal to the sum of the zero-input and zero-state responses [i.e., y(n) = yzi (n) + yzs (n)]. 2. The principle of superposition applies to the zero-state response (zero-state linear). 3. The principle of superposition applies to the zero-input response (zero-input linear). A system that does not satisfy all three separate requirements is by definition nonlinear. Obviously, for a relaxed system, yzi (n) = 0, and thus requirement 2, which is the definition of linearity given in Section 2.4, is sufficient. We illustrate the application of these requirements by a simple example. EXAMPLE 4.2 Determine if the recursive system defined by the difference equation y(n) = ay(n − 1) + x(n) is linear. Solution.

By combining (4.9) and (4.11), we obtain (4.8), which can be expressed as y(n) = yzi (n) + yzs (n)

Thus the first requirement for linearity is satisfied. To check for the second requirement, let us assume that x(n) = c1 x1 (n) + c2 x2 (n). Then (4.9) gives yzs (n) =

n 

a k [c1 x1 (n − k) + c2 x2 (n − k)]

k=0

= c1

n 

a k x1 (n − k) + c2

k=0

n 

a k x2 (n − k)

k=0

(1) (2) = c1 yzs (n) + c2 yzs (n)

Hence yzs (n) satisfies the principle of superposition, and thus the system is zero-state linear.

98

Discrete-Time Signals and Systems

Now let us assume that y(−1) = c1 y1 (−1) + c2 y2 (−1). From (4.11) we obtain yzi (n) = a n+1 [c1 y1 (−1) + c2 y2 (−1)] = c1 a n+1 y1 (−1) + c2 a n+1 y2 (−1) = c1 yzi(1) (n) + c2 yzi(2) (n) Hence the system is zero-input linear. Since the system satisfies all three conditions for linearity, it is linear.

Although it is somewhat tedious, the procedure used in Example 4.2 to demonstrate linearity for the system described by the first-order difference equation carries over directly to the general recursive systems described by the constant-coefficient difference equation given in (4.13). Hence, a recursive system described by the linear difference equation in (4.13) also satisfies all three conditions in the definition of linearity, and therefore it is linear. The next question that arises is whether or not the causal linear system described by the linear constant-coefficient difference equation in (4.13) is time invariant. This is fairly easy, when dealing with systems described by explicit input–output mathematical relationships. Clearly, the system described by (4.13) is time invariant because the coefficients ak and bk are constants. On the other hand, if one or more of these coefficients depends on time, the system is time variant, since its properties change as a function of time. Thus we conclude that the recursive system described by a linear constant-coefficient difference equation is linear and time invariant. The final issue is the stability of the recursive system described by the linear, constant-coefficient difference equation in (4.13). In Section 3.6 we introduced the concept of bounded input–bounded output (BIBO) stability for relaxed systems. For nonrelaxed systems that may be nonlinear, BIBO stability should be viewed with some care. However, in the case of a linear time-invariant recursive system described by the linear constant-coefficient difference equation in (4.13), it suffices to state that such a system is BIBO stable if and only if for every bounded input and every bounded initial condition, the total system response is bounded. EXAMPLE 4.3 Determine if the linear time-invariant recursive system described by the difference equation given in (4.7) is stable. Solution. Let us assume that the input signal x(n) is bounded in amplitude, that is, |x(n)| ≤ Mx < ∞ for all n ≥ 0. From (4.8) we have  n      n+1 k a x(n − k) , n≥0 |y(n)| ≤ |a y(−1)| +    k=0

≤ |a|n+1 |y(−1)| + Mx

n 

|a|k ,

n≥0

k=0

≤ |a|n+1 |y(−1)| + Mx

1 − |a|n+1 = My , 1 − |a|

n≥0

99

Discrete-Time Signals and Systems

If n is finite, the bound My is finite and the output is bounded independently of the value of a . However, as n → ∞, the bound My remains finite only if |a| < 1 because |a|n → 0 as n → ∞. Then My = Mx /(1 − |a|). Thus the system is stable only if |a| < 1.

For the simple first-order system in Example 4.3, we were able to express the condition for BIBO stability in terms of the system parameter a , namely |a| < 1. We should stress, however, that this task becomes more difficult for higher-order systems. Fortunately, other simple and more efficient techniques exist for investigating the stability of recursive systems.

4.3

Solution of Linear Constant-Coefficient Difference Equations

Given a linear constant-coefficient difference equation as the input–output relationship describing a linear time-invariant system, our objective in this subsection is to determine an explicit expression for the output y(n). The method that is developed is termed the direct method. An alternative method based on the z-transform is beyond the scope of this chapter. The z -transform approach is called the indirect method. Basically, the goal is to determine the output y(n), n ≥ 0, of the system given a specific input x(n), n ≥ 0, and a set of initial conditions. The direct solution method assumes that the total solution is the sum of two parts: y(n) = yh (n) + yp (n) The part yh (n) is known as the homogeneous or complementary solution, whereas yp (n) is called the particular solution. The homogeneous solution of a difference equation. We begin the problem of solving the linear constant-coefficient difference equation given by (4.13) by obtaining first the solution to the homogeneous difference equation N 

ak y(n − k) = 0

(4.14)

k=0

The procedure for solving a linear constant-coefficient difference equation directly is very similar to the procedure for solving a linear constant-coefficient differential equation. Basically, we assume that the solution is in the form of an exponential, that is, yh (n) = λn (4.15) where the subscript h on y(n) is used to denote the solution to the homogeneous difference equation. If we substitute this assumed solution in (4.14), we obtain the polynomial equation N  ak λn−k = 0 k=0

100

Discrete-Time Signals and Systems

or λn−N (λN + a1 λN−1 + a2 λN−2 + · · · + aN−1 λ + aN ) = 0

(4.16)

The polynomial in parentheses is called the characteristic polynomial of the system. In general, it has N roots, which we denote as λ1 , λ2 , . . . , λN . The roots can be real or complex valued. In practice the coefficients a1 , a2 , . . . , aN are usually real. Complex-valued roots occur as complex-conjugate pairs. Some of the N roots may be identical, in which case we have multiple-order roots. For the moment, let us assume that the roots are distinct, that is, there are no multiple-order roots. Then the most general solution to the homogeneous difference equation in (4.14) is yh (n) = C1 λn1 + C2 λn2 + · · · + CN λnN

(4.17)

where C1 , C2 , . . . , CN are weighting coefficients. These coefficients are determined from the initial conditions specified for the system. Since the input x(n) = 0, (4.17) can be used to obtain the zero-input response of the system. The following examples illustrate the procedure. EXAMPLE 4.4 Determine the homogeneous solution of the system described by the first-order difference equation (4.18) y(n) + a1 y(n − 1) = x(n) Solution.

The assumed solution obtained by setting x(n) = 0 is yh (n) = λn

When we substitute this solution in (4.18), we obtain [with x(n) = 0] λn + a1 λn−1 = 0 λn−1 (λ + a1 ) = 0 λ = −a1 Therefore, the solution to the homogeneous difference equation is yh (n) = Cλn = C(−a1 )n

(4.19)

The zero-input response of the system can be determined from (4.18) and (4.19). With x(n) = 0, (4.18) yields y(0) = −a1 y(−1) On the other hand, from (4.19) we have yh (0) = C and hence the zero-input response of the system is yzi (n) = (−a1 )n+1 y(−1),

n≥0

(4.20)

With a = −a1 , this result is consistent with (4.11) for the first-order system, which was obtained earlier by iteration of the difference equation.

101

Discrete-Time Signals and Systems

EXAMPLE 4.5 Determine the zero-input response of the system described by the homogeneous second-order difference equation y(n) − 3y(n − 1) − 4y(n − 2) = 0 (21) Solution. First we determine the solution to the homogeneous equation. We assume the solution to be the exponential yh (n) = λn Upon substitution of this solution into (4.21), we obtain the characteristic equation λn − 3λn−1 − 4λn−2 = 0 λn−2 (λ2 − 3λ − 4) = 0 Therefore, the roots are λ = −1, 4, and the general form of the solution to the homogeneous equation is yh (n) = C1 λn1 + C2 λn2 (4.22) = C1 (−1)n + C2 (4)n The zero-input response of the system can be obtained from the homogenous solution by evaluating the constants in (4.22), given the initial conditions y(−1) and y(−2). From the difference equation in (4.21) we have y(0) = 3y(−1) + 4y(−2) y(1) = 3y(0) + 4y(−1) = 3[3y(−1) + 4y(−2)] + 4y(−1) = 13y(−1) + 12y(−2) On the other hand, from (4.22) we obtain y(0) = C1 + C2 y(1) = −C1 + 4C2 By equating these two sets of relations, we have C1 + C2 = 3y(−1) + 4y(−2) −C1 + 4C2 = 13y(−1) + 12y(−2) The solution of these two equations is 4 1 C1 = − y(−1) + y(−2) 5 5 C2 =

102

16 16 y(−1) + y(−2) 5 5

Discrete-Time Signals and Systems

Therefore, the zero-input response of the system is 4 1 yzi (n) = [− y(−1) + y(−2)](−1)n 5 5

(4.23)

16 16 + [ y(−1) + y(−2)](4)n , 5 5

n≥0

For example, if y(−2) = 0 and y(−1) = 5, then C1 = −1, C2 = 16, and hence yzi (n) = (−1)n+1 + (4)n+2 ,

n≥0

These examples illustrate the method for obtaining the homogeneous solution and the zero-input response of the system when the characteristic equation contains distinct roots. On the other hand, if the characteristic equation contains multiple roots, the form of the solution given in (4.17) must be modified. For example, if λ1 is a root of multiplicity m, then (4.17) becomes yh (n) = C1 λn1 + C2 nλn1 + C3 n2 λn1 + · · · + Cm nm−1 λn1 + Cm+1 λnm+1 + · · · + CN λn

(4.24)

The particular solution yp (n) is required to satisfy the difference equation (4.13) for the specific input signal x(n), n ≥ 0. In other words, yp (n) is any solution satisfying The particular solution of the difference equation.

N  k=0

ak yp (n − k) =

M 

bk x(n − k),

a0 = 1

(4.25)

k=0

To solve (4.25), we assume for yp (n), a form that depends on the form of the input x(n). The following example illustrates the procedure. EXAMPLE 4.6 Determine the particular solution of the first-order difference equation y(n) + a1 y(n − 1) = x(n),

|a1 | < 1

(4.26)

when the input x(n) is a unit step sequence, that is, x(n) = u(n) Solution. Since the input sequence x(n) is a constant for n ≥ 0, the form of the solution that we assume is also a constant. Hence the assumed solution of the difference equation to the forcing function x(n), called the particular solution of the difference equation, is yp (n) = Ku(n)

103

Discrete-Time Signals and Systems

where K is a scale factor determined so that (4.26) is satisfied. Upon substitution of this assumed solution into (4.26), we obtain Ku(n) + a1 Ku(n − 1) = u(n) To determine K, we must evaluate this equation for any n ≥ 1, where none of the terms vanish. Thus K + a1 K = 1 K=

1 1 + a1

Therefore, the particular solution to the difference equation is yp (n) =

1 u(n) 1 + a1

(4.27)

In this example, the input x(n), n ≥ 0, is a constant and the form assumed for the particular solution is also a constant. If x(n) is an exponential, we would assume that the particular solution is also an exponential. If x(n) were a sinusoid, then yp (n) would also be a sinusoid. Thus our assumed form for the particular solution takes the basic form of the signal x(n). Table 1 provides the general form of the particular solution for several types of excitation. EXAMPLE 4.7 Determine the particular solution of the difference equation y(n) =

5 1 y(n − 1) − y(n − 2) + x(n) 6 6

when the forcing function x(n) = 2n , n ≥ 0 and zero elsewhere.

General Form of the Particular Solution for Several Types of Input Signals Input Signal, Particular Solution, x(n) yp (n) A (constant) K AM n KM n AnM K0 nM + K1 nM−1 + · · · + KM n M n A n A (K0 nM + K1 nM−1 + · · · + KM )  A cos ω0 n K1 cos ω0 n + K2 sin ω0 n A sin ω0 n

TABLE 1

104

Discrete-Time Signals and Systems

Solution.

The form of the particular solution is yp (n) = K2n ,

n≥0

Upon substitution of yp (n) into the difference equation, we obtain K2n u(n) =

5 1 K2n−1 u(n − 1) − K2n−2 u(n − 2) + 2n u(n) 6 6

To determine the value of K , we can evaluate this equation for any n ≥ 2, where none of the terms vanish. Thus we obtain 1 5 4K = (2K) − K + 4 6 6 and hence K = 85 . Therefore, the particular solution is yp (n) =

8 n 2 , 5

n≥0

We have now demonstrated how to determine the two components of the solution to a difference equation with constant coefficients. These two components are the homogeneous solution and the particular solution. From these two components, we construct the total solution from which we can obtain the zero-state response. The linearity property of the linear constant-coefficient difference equation allows us to add the homogeneous solution and the particular solution in order to obtain the total solution. Thus

The total solution of the difference equation.

y(n) = yh (n) + yp (n) The resultant sum y(n) contains the constant parameters {Ci } embodied in the homogeneous solution component yh (n). These constants can be determined to satisfy the initial conditions. The following example illustrates the procedure. EXAMPLE 4.8 Determine the total solution y(n), n ≥ 0, to the difference equation y(n) + a1 y(n − 1) = x(n)

(4.28)

when x(n) is a unit step sequence [i.e., x(n) = u(n)] and y(−1) is the initial condition. Solution.

From (4.19) of Example 4.4, the homogeneous solution is yh (n) = C(−a1 )n

and from (4.26) of Example 4.6, the particular solution is yp (n) =

1 u(n) 1 + a1

105

Discrete-Time Signals and Systems

Consequently, the total solution is y(n) = C(−a1 )n +

1 , 1 + a1

n≥0

(4.29)

where the constant C is determined to satisfy the initial condition y(−1). In particular, suppose that we wish to obtain the zero-state response of the system described by the first-order difference equation in (4.28). Then we set y(−1) = 0. To evaluate C , we evaluate (4.28) at n = 0, obtaining y(0) + a1 y(−1) = 1 Hence, y(0) = 1 − a1 y(−1) On the other hand, (4.29) evaluated at n = 0 yields y(0) = C +

1 1 + a1

By equating these two relations, we obtain C+

1 = −a1 y(−1) + 1 1 + a1 C = −a1 y(−1) +

a1 1 + a1

Finally, if we substitute this value of C into (4.29), we obtain y(n) = (−a1 )n+1 y(−1) +

1 − (−a1 )n+1 , 1 + a1

n≥0

(4.30)

= yzi (n) + yzs (n)

We observe that the system response as given by (4.30) is consistent with the response y(n) given in (4.8) for the first-order system (with a = −a) 1, which was obtained by solving the difference equation iteratively. Furthermore, we note that the value of the constant C depends both on the initial condition y(−1) and on the excitation function. Consequently, the value of C influences both the zero-input response and the zero-state response. We further observe that the particular solution to the difference equation can be obtained from the zero-state response of the system. Indeed, if |a1 | < 1, which is the condition for stability of the system, as will be shown in Section 4.4, the limiting value of yzs (n) as n approaches infinity is the particular solution, that is, yp (n) = lim yzs (n) = n→∞

1 1 + a1

Since this component of the system response does not go to zero as n approaches infinity, it is usually called the steady-state response of the system. This response persists as long as the input persists. The component that dies out as n approaches infinity is called the transient response of the system. The following example illustrates the evaluation of the total solution for a secondorder recursive system.

106

Discrete-Time Signals and Systems

EXAMPLE 4.9 Determine the response y(n), n ≥ 0, of the system described by the second-order difference equation y(n) − 3y(n − 1) − 4y(n − 2) = x(n) + 2x(n − 1) (4.31) when the input sequence is x(n) = 4n u(n) Solution. We have already determined the solution to the homogeneous difference equation for this system in Example 4.5. From (4.22) we have yh (n) = C1 (−1)n + C2 (4)n

(4.32)

The particular solution to (4.31) is assumed to be an exponential sequence of the same form as x(n). Normally, we could assume a solution of the form yp (n) = K(4)n u(n) However, we observe that yp (n) is already contained in the homogeneous solution, so that this particular solution is redundant. Instead, we select the particular solution to be linearly independent of the terms contained in the homogeneous solution. In fact, we treat this situation in the same manner as we have already treated multiple roots in the characteristic equation. Thus we assume that (4.33) yp (n) = Kn(4)n u(n) Upon substitution of (4.33) into (4.31), we obtain Kn(4)n u(n) − 3K(n − 1)(4)n−1 u(n − 1) − 4K(n − 2)(4)n−2 u(n − 2) = (4)n u(n) + 2(4)n−1 u(n − 1) To determine K , we evaluate this equation for any n ≥ 2, where none of the unit step terms vanish. To simplify the arithmetic, we select n = 2, from which we obtain K = 65 . Therefore, 6 (4.34) yp (n) = n(4)n u(n) 5 The total solution to the difference equation is obtained by adding (4.32) to (4.34). Thus 6 (4.35) y(n) = C1 (−1)n + C2 (4)n + n(4)n , n≥0 5 where the constants C1 and C2 are determined such that the initial conditions are satisfied. To accomplish this, we return to (4.31), from which we obtain y(0) = 3y(−1) + 4y(−2) + 1 y(1) = 3y(0) + 4y(−1) + 6 = 13y(−1) + 12y(−2) + 9 On the other hand, (4.35) evaluated at n = 0 and n = 1 yields y(0) = C1 + C2 y(1) = −C1 + 4C2 +

24 5

107

Discrete-Time Signals and Systems

We can now equate these two sets of relations to obtain C1 and C2 . In so doing, we have the response due to initial conditions y(−1) and y(−2) (the zero-input response), and the zero-state response. Since we have already solved for the zero-input response in Example 4.5, we can simplify the computations above by setting y(−1) = y(−2) = 0. Then we have C1 + C2 = 1 −C1 + 4C2 +

24 =9 5

1 Hence C1 = − 25 and C2 = 26 . Finally, we have the zero-state response to the forcing function 25 n x(n) = (4) u(n) in the form

yzs (n) = −

1 26 6 (−1)n + (4)n + n(4)n , 25 25 5

n≥0

(4.36)

The total response of the system, which includes the response to arbitrary initial conditions, is the sum of (4.23) and (4.36).

4.4

The Impulse Response of a Linear Time-Invariant Recursive System

The impulse response of a linear time-invariant system was previously defined as the response of the system to a unit sample excitation [i.e., x(n) = δ(n)]. In the case of a recursive system, h(n) is simply equal to the zero-state response of the system when the input x(n) = δ(n) and the system is initially relaxed. For example, in the simple first-order recursive system given in (4.7), the zerostate response given in (4.8), is yzs (n) =

n 

a k x(n − k)

(4.37)

k=0

When x(n) = δ(n) is substituted into (4.37), we obtain yzs (n) =

n 

a k δ(n − k)

k=0

= an,

n≥0

Hence the impulse response of the first-order recursive system described by (4.7) is h(n) = a n u(n)

(4.38)

as indicated in Section 4.2. In the general case of an arbitrary, linear time-invariant recursive system, the zero-state response expressed in terms of the convolution summation is yzs (n) =

n  k=0

108

h(k)x(n − k),

n≥0

(4.39)

Discrete-Time Signals and Systems

When the input is an impulse [i.e., x(n) = δ(n)], (4.39) reduces to yzs (n) = h(n)

(4.40)

Now, let us consider the problem of determining the impulse response h(n) given a linear constant-coefficient difference equation description of the system. In terms of our discussion in the preceding subsection, we have established the fact that the total response of the system to any excitation function consists of the sum of two solutions of the difference equation: the solution to the homogeneous equation plus the particular solution to the excitation function. In the case where the excitation is an impulse, the particular solution is zero, since x(n) = 0 for n > 0, that is, yp (n) = 0 Consequently, the response of the system to an impulse consists only of the solution to the homogeneous equation, with the {Ck } parameters evaluated to satisfy the initial conditions dictated by the impulse. The following example illustrates the procedure for obtaining h(n) given the difference equation for the system. EXAMPLE 4.10 Determine the impulse response h(n) for the system described by the second-order difference equation (4.41) y(n) − 3y(n − 1) − 4y(n − 2) = x(n) + 2x(n − 1) Solution. We have already determined in Example 4.5 that the solution to the homogeneous difference equation for this system is yh (n) = C1 (−1)n + C2 (4)n ,

n≥0

(4.42)

Since the particular solution is zero when x(n) = δ(n), the impulse response of the system is simply given by (4.42), where C 1 and C2 must be evaluated to satisfy (4.41). For n = 0 and n = 1, (4.41) yields y(0) = 1 y(1) = 3y(0) + 2 = 5 where we have imposed the conditions y(−1) = y(−2) = 0, since the system must be relaxed. On the other hand, (4.42) evaluated at n = 0 and n = 1 yields y(0) = C1 + C2 y(1) = −C1 + 4C2 By solving these two sets of equations for C1 and C2 , we obtain 1 C1 = − , 5

C2 =

6 5

Therefore, the impulse response of the system is

1 6 h(n) = − (−1)n + (4)n u(n) 5 5

109

Discrete-Time Signals and Systems

When the system is described by an Nth-order linear difference equation of the type given in (4.13), the solution of the homogeneous equation is yh (n) =

N 

Ck λ n k

k=1

when the roots {λk } of the characteristic polynomial are distinct. Hence the impulse response of the system is identical in form, that is, h(n) =

N 

Ck λnk

(4.43)

k=1

where the parameters {Ck } are determined by setting the initial conditions y(−1) = · · · = y(−N ) = 0. This form of h(n) allows us to easily relate the stability of a system, described by an Nth-order difference equation, to the values of the roots of the characteristic polynomial. Indeed, since BIBO stability requires that the impulse response be absolutely summable, then, for a causal system, we have ∞  n=0

 N  ∞  N ∞      n  |h(n)| = C λ k ≤ |C | |λk |n   k k  

Now if |λk | < 1 for all k, then

n=0 k=1

∞ 

k=1

n=0

|λk |n < ∞

n=0

and hence

∞ 

|h(n)| < ∞

n=0

On the other hand, if one or more of the |λk | ≥ 1, h(n) is no longer absolutely summable, and consequently, the system is unstable. Therefore, a necessary and sufficient condition for the stability of a causal IIR system described by a linear constant-coefficient difference equation is that all roots of the characteristic polynomial be less than unity in magnitude. The reader may verify that this condition carries over to the case where the system has roots of multiplicity m. Finally we note that any recursive system described by a linear constant-coefficient difference equation is an IIR system. The converse is not true, however. That is, not every linear time-invariant IIR system can be described by a linear constantcoefficient difference equation. In other words, recursive systems described by linear constant-coefficient difference equations are a subclass of linear time-invariant IIR systems.

110

Discrete-Time Signals and Systems

5

Implementation of Discrete-Time Systems Our treatment of discrete-time systems has been focused on the time-domain characterization and analysis of linear time-invariant systems described by constantcoefficient linear difference equations. Additional analytical methods are developed, where we characterize and analyze LTI systems in the frequency domain. Two oth er important topics that will be treated later are the design and implementation of these systems. In practice, system design and implementation are usually treated jointly rather than separately. Often, the system design is driven by the method of implementation and by implementation constraints, such as cost, hardware limitations, size limitations, and power requirements. At this point, we have not as yet developed the necessary analysis and design tools to treat such complex issues. However, we have developed sufficient background to consider some basic implementation methods for realizations of LTI systems described by linear constant-coefficient difference equations.

5.1

Structures for the Realization of Linear Time-Invariant Systems

In this subsection we describe structures for the realization of systems described by linear constant-coefficient difference equations. As a beginning, let us consider the first-order system y(n) = −a1 y(n − 1) + b0 x(n) + b1 x(n − 1)

(5.1)

which is realized as in Fig. 5.1(a). This realization uses separate delays (memory) for both the input and output signal samples and it is called a direct form I structure. Note that this system can be viewed as two linear time-invariant systems in cascade. The first is a nonrecursive system described by the equation v(n) = b0 x(n) + b1 x(n − 1)

(5.2)

whereas the second is a recursive system described by the equation y(n) = −a1 y(n − 1) + v(n)

(5.3)

However, as we have seen in Section 3.4, if we interchange the order of the cascaded linear time-invariant systems, the overall system response remains the same. Thus if we interchange the order of the recursive and nonrecursive systems, we obtain an alternative structure for the realization of the system described by (5.1). The resulting system is shown in Fig. 5.1(b). From this figure we obtain the two difference equations w(n) = −a1 w(n − 1) + x(n)

(5.4)

y(n) = b0 w(n) + b1 w(n − 1)

(5.5)

111

Discrete-Time Signals and Systems

x(n)

b0

+

ν(n)

y(n)

+

z−1

z−1 −a1

b1 (a) ω(n)

x(n) + z−1

b0

y(n) +

z−1

−a1

b1 ω(n − 1)

ω(n − 1) (b)

x(n)

ω(n)

+

b0

+

y(n)

z−1

Figure 5.1

−a1

Steps in converting from the direct form I realization in (a) to the direct form II realization in (c).

b1 ω(n − 1) (c)

which provide an alternative algorithm for computing the output of the system described by the single difference equation given in (5.1). In other words, the two difference equations (5.4) and (5.5) are equivalent to the single difference equation (5.1). A close observation of Fig. 5.1 reveals that the two delay elements contain the same input w(n) and hence the same output w(n − 1). Consequently, these two elements can be merged into one delay, as shown in Fig. 5.1(c). In contrast to the direct form I structure, this new realization requires only one delay for the auxiliary quantity w(n), and hence it is more efficient in terms of memory requirements. It is called the direct form II structure and it is used extensively in practical applications. These structures can readily be generalized for the general linear time-invariant recursive system described by the difference equation

y(n) = −

N  k=1

ak y(n − k) +

M 

bk x(n − k)

(5.6)

k=0

Figure 5.2 illustrates the direct form I structure for this system. This structure requires M + N delays and N + M + 1 multiplications. It can be viewed as the

112

Discrete-Time Signals and Systems

cascade of a nonrecursive system v(n) =

M 

bk x(n − k)

(5.7)

k=0

and a recursive system y(n) = −

N 

ak y(n − k) + v(n)

(5.8)

k=1

By reversing the order of these two systems, as was previously done for the firstorder system, we obtain the direct form II structure shown in Fig. 5.3 for N > M . This structure is the cascade of a recursive system w(n) = −

N 

ak w(n − k) + x(n)

(5.9)

k=1

followed by a nonrecursive system y(n) =

M 

bk w(n − k)

(5.10)

k=0

x(n)

b0

+

ν(n)

y(n)

+

z−1

z−1 b1

+

+

−a1

bM − 1

+

+

+

+

−a2





b2



z−1



z−1

−aN − 1

z−1

z−1 bM

Figure 5.2

−aN

Direct form I structure of the system described by (5.6).

113

Discrete-Time Signals and Systems

We observe that if N ≥ M , this structure requires a number of delays equal to the order N of the system. However, if M > N , the required memory is specified by M . Figure 5.3 can easily be modified to handle this case. Thus the direct form II structure requires M + N + 1 multiplications and max{M, N } delays. Because it requires the minimum number of delays for the realization of the system described by (5.6), it is sometimes called a canonic form. A special case of (5.6) occurs if we set the system parameters a k = 0, k = 1, . . . , N . Then the input–output relationship for the system reduces to y(n) =

M 

bk x(n − k)

(5.11)

k=0

which is a nonrecursive linear time-invariant system. This system views only the most recent M + 1 input signal samples and, prior to addition, weights each sample by the appropriate coefficient bk from the set {bk }. In other words, the system output is basically a weighted moving average of the input signal. For this reason it is sometimes called a moving average (MA) system. Such a system is an FIR system x(n)

ω(n) +

b0

+

z−1 +

−a1

ω(n − 1)

b1

+

z−1 +

−a2

ω(n − 2)

b2

+

z−1

+

b3

−aN−2

ω(n − M)

bM

+ …

ω(n − 3)



−a3



+

(M = N − 2) z−1

+

−aN−1 z−1 −aN

ω(n − N)

Figure 5.3 Direct form II structure for the system described by (5.6).

114

y(n)

Discrete-Time Signals and Systems

with an impulse response h(k) equal to the coefficients bk , that is,  bk , 0 ≤ k ≤ M h(k) = 0, otherwise

(5.12)

If we return to (5.6) and set M = 0, the general linear time-invariant system reduces to a “purely recursive” system described by the difference equation y(n) = −

N 

ak y(n − k) + b0 x(n)

(5.13)

k=1

In this case the system output is a weighted linear combination of N past outputs and the present input. Linear time-invariant systems described by a second-order difference equation are an important subclass of the more general systems described by (5.6) or (5.10) or (5.13). The reason for their importance will be explained later when we discuss quantization effects. Suffice to say at this point that second-order systems are usually used as basic building blocks for realizing higher-order systems. The most general second-order system is described by the difference equation y(n) = − a1 y(n − 1) − a2 y(n − 2) + b0 x(n) + b1 x(n − 1) + b2 x(n − 2)

(5.14)

which is obtained from ( 5.6) by setting N = 2 and M = 2. The direct form II structure for realizing this system is shown in Fig. 5.4(a). If we set a1 = a2 = 0, then (5.14) reduces to y(n) = b0 x(n) + b1 x(n − 1) + b2 x(n − 2)

(5.15)

which is a special case of the FIR system described by (5.11). The structure for realizing this system is shown in Fig. 5.4(b). Finally, if we set b 1 = b2 = 0 in (5.14), we obtain the purely recursive second-order system described by the difference equation y(n) = −a1 y(n − 1) − a2 y(n − 2) + b0 x(n)

(5.16)

which is a special case of (5.13). The structure for realizing this system is shown in Fig. 5.4(c).

5.2

Recursive and Nonrecursive Realizations of FIR Systems

We have already made the distinction between FIR and IIR systems, based on whether the impulse response h(n) of the system has a finite duration, or an infinite duration. We have also made the distinction between recursive and nonrecursive systems. Basically, a causal recursive system is described by an input–output equation of the form y(n) = F [y(n − 1), . . . , y(n − N), x(n), . . . , x(n − M)]

(5.17)

115

Discrete-Time Signals and Systems

x(n)

b0

+

+

y(n)

z −1 −a1

+

b1

+

z−1 −a2

b2 (a)

x(n)

z−1

z−1

b0

b1

b2

+

+

y(n)

(b) x(n) b0

+

+ −a1

y(n) −a2

z−1

z−1 (c)

Structures for the realization of second-order systems: (a) general second-order system; (b) FIR system; (c) “purely recursive system.”

Figure 5.4

and for a linear time-invariant system specifically, by the difference equation y(n) = −

N  k=1

ak y(n − k) +

M 

bk x(n − k)

(5.18)

k=0

On the other hand, causal nonrecursive systems do not depend on past values of the output and hence are described by an input–output equation of the form y(n) = F [x(n), x(n − 1), . . . , x(n − M)]

(5.19)

and for linear time-invariant systems specifically, by the difference equation in (5.18) with ak = 0 for k = 1, 2, . . . , N . In the case of FIR systems, we have already observed that it is always possible to realize such systems nonrecursively. In fact, with ak = 0, k = 1, 2, . . . , N , in (5.18),

116

Discrete-Time Signals and Systems

x(n)

z−1

z−1

+

z−1

z−1



y(n)



+

+ 1 M+1

Figure 5.5 Nonrecursive realization of an FIR moving average system.

we have a system with an input–output equation

y(n) =

M 

bk x(n − k)

(5.20)

k=0

This is a nonrecursive and FIR system. As indicated in (5.12), the impulse response of the system is simply equal to the coefficients {bk }. Hence every FIR system can be realized nonrecursively. On the other hand, any FIR system can also be realized recursively. Although the general proof of this statement is given later, we shall give a simple example to illustrate the point. Suppose that we have an FIR system of the form 1  x(n − k) M +1 M

y(n) =

(5.21)

k=0

for computing the moving average of a signal x(n). Clearly, this system is FIR with impulse response 1 h(n) = , 0≤n≤M M +1 Figure 5.5 illustrates the structure of the nonrecursive realization of the system. Now, suppose that we express (5.21) as 1  x(n − 1 − k) M +1 M

y(n) =

k=0

+

1 [x(n) − x(n − 1 − M)] M +1

= y(n − 1) +

(5.22)

1 [x(n) − x(n − 1 − M)] M +1

Now, (5.22) represents a recursive realization of the FIR system. The structure of this recursive realization of the moving average system is illustrated in Fig. 5.6.

117

Discrete-Time Signals and Systems

x(n)

z−1

z−1



z−1

x(n − M − 1)

− +

y(n) +

+ 1 M+1

z−1

y(n − 1)

Figure 5.6 Recursive realization of an FIR moving average system.

In summary, we can think of the terms FIR and IIR as general characteristics that distinguish a type of linear time-invariant system, and of the terms recursive and nonrecursive as descriptions of the structures for realizing or implementing the system.

6

Correlation of Discrete-Time Signals A mathematical operation that closely resembles convolution is correlation. Just as in the case of convolution, two signal sequences are involved in correlation. In contrast to convolution, however, our objective in computing the correlation between the two signals is to measure the degree to which the two signals are similar and thus to extract some information that depends to a large extent on the application. Correlation of signals is often encountered in radar, sonar, digital communications, geology, and other areas in science and engineering. To be specific, let us suppose that we have two signal sequences x(n) and y(n) that we wish to compare. In radar and active sonar applications, x(n) can represent the sampled version of the transmitted signal and y(n) can represent the sampled version of the received signal at the output of the analog-to-digital (A/D) converter. If a target is present in the space being searched by the radar or sonar, the received signal y(n) consists of a delayed version of the transmitted signal, reflected from the target, and corrupted by additive noise. Figure 6.1 depicts the radar signal reception problem. We can represent the received signal sequence as y(n) = αx(n − D) + w(n)

(6.1)

where α is some attenuation factor representing the signal loss involved in the roundtrip transmission of the signal x(n), D is the round-trip delay, which is assumed to be an integer multiple of the sampling interval, and w(n) represents the additive noise that is picked up by the antenna and any noise generated by the electronic components and amplifiers contained in the front end of the receiver. On the other hand, if there is no target in the space searched by the radar and sonar, the received signal y(n) consists of noise alone.

118

Discrete-Time Signals and Systems

al

ted mit

ns

Tra

Figure 6.1

nal

sign

ed ect

sig

fl

Re

Radar target detection.

Having the two signal sequences, x(n), which is called the reference signal or transmitted signal, and y(n), the received signal, the problem in radar and sonar detection is to compare y(n) and x(n) to determine if a target is present and, if so, to determine the time delay D and compute the distance to the target. In practice, the signal x(n − D) is heavily corrupted by the additive noise to the point where a visual inspection of y(n) does not reveal the presence or absence of the desired signal reflected from the target. Correlation provides us with a means for extracting this important information from y(n). Digital communications is another area where correlation is often used. In digital communications the information to be transmitted from one point to another is usually converted to binary form, that is, a sequence of zeros and ones, which are then transmitted to the intended receiver. To transmit a 0 we can transmit the signal sequence x0 (n) for 0 ≤ n ≤ L − 1, and to transmit a 1 we can transmit the signal sequence x1 (n) for 0 ≤ n ≤ L−1, where L is some integer that denotes the number of samples in each of the two sequences. Very often, x1 (n) is selected to be the negative of x0 (n). The signal received by the intended receiver may be represented as y(n) = xi (n) + w(n),

i = 0, 1,

0≤n≤L−1

(6.2)

where now the uncertainty is whether x0 (n) or x1 (n) is the signal component in y(n), and w(n) represents the additive noise and other interference inherent in any communication system. Again, such noise has its origin in the electronic components contained in the front end of the receiver. In any case, the receiver knows the possible transmitted sequences x0 (n) and x1 (n) and is faced with the task of comparing the received signal y(n) with both x0 (n) and x1 (n) to determine which of the two signals better matches y(n). This comparison process is performed by means of the correlation operation described in the following subsection.

119

Discrete-Time Signals and Systems

6.1

Crosscorrelation and Autocorrelation Sequences

Suppose that we have two real signal sequences x(n) and y(n) each of which has finite energy. The crosscorrelation of x(n) and y(n) is a sequence rxy (l), which is defined as ∞  x(n)y(n − l), l = 0, ±1, ±2, . . . (6.3) rxy (l) = n=−∞

or, equivalently, as rxy (l) =

∞ 

x(n + l)y(n),

l = 0, ±1, ±2, . . .

(6.4)

n=−∞

The index l is the (time) shift (or lag) parameter and the subscripts xy on the crosscorrelation sequence rxy (l) indicate the sequences being correlated. The order of the subscripts, with x preceding y , indicates the direction in which one sequence is shifted, relative to the other. To elaborate, in (6.3), the sequence x(n) is left unshifted and y(n) is shifted by l units in time, to the right for l positive and to the left for l negative. Equivalently, in (6.4), the sequence y(n) is left unshifted and x(n) is shifted by l units in time, to the left for l positive and to the right for l negative. But shifting x(n) to the left by l units relative to y(n) is equivalent to shifting y(n) to the right by l units relative to x(n). Hence the computations (6.3) and (6.4) yield identical crosscorrelation sequences. If we reverse the roles of x(n) and y(n) in (6.3) and (6.4) and therefore reverse the order of the indices xy , we obtain the crosscorrelation sequence ryx (l) =

∞ 

y(n)x(n − l)

(6.5)

y(n + l)x(n)

(6.6)

n=−∞

or, equivalently, ryx (l) =

∞  n=−∞

By comparing (6.3) with (6.6) or (6.4) with (6.5), we conclude that rxy (l) = ryx (−l)

(6.7)

Therefore, ryx (l) is simply the folded version of rxy (l), where the folding is done with respect to l = 0. Hence, ryx (l) provides exactly the same information as rxy (l), with respect to the similarity of x(n) to y(n). EXAMPLE 6.1 Determine the crosscorrelation sequence rxy (l) of the sequences x(n) = {. . . , 0, 0, 2, −1, 3, 7, 1, 2, −3, 0, 0, . . .} ↑

y(n) = {. . . , 0, 0, 1, −1, 2, −2, 4, 1, −2, 5, 0, 0, . . .} ↑

120

Discrete-Time Signals and Systems

Solution.

Let us use the definition in (6.3) to compute rxy (l). For l = 0 we have rxy (0) =

∞ 

x(n)y(n)

n=−∞

The product sequence v0 (n) = x(n)y(n) is v0 (n) = {. . . , 0, 0, 2, 1, 6, −14, 4, 2, 6, 0, 0, . . .} ↑

and hence the sum over all values of n is rxy (0) = 7 For l > 0, we simply shift y(n) to the right relative to x(n) by l units, compute the product sequence vl (n) = x(n)y(n − l), and finally, sum over all values of the product sequence. Thus we obtain rxy (1) = 13, rxy (2) = −18, rxy (3) = 16, rxy (4) = −7 rxy (5) = 5, rxy (6) = −3, rxy (l) = 0, l≥7 For l < 0, we shift y(n) to the left relative to x(n) by l units, compute the product sequence vl (n) = x(n)y(n − l), and sum over all values of the product sequence. Thus we obtain the values of the crosscorrelation sequence rxy (−1) = 0, rxy (−5) = 19,

rxy (−2) = 33, rxy (−6) = −9,

rxy (−3) = −14, rxy (−7) = 10,

rxy (−4) = 36 rxy (l) = 0, l ≤ −8

Therefore, the crosscorrelation sequence of x(n) and y(n) is rxy (l) = {10, −9, 19, 36, −14, 33, 0, 7, 13, −18, 16, −7, 5, −3} ↑

The similarities between the computation of the crosscorrelation of two sequences and the convolution of two sequences is apparent. In the computation of convolution, one of the sequences is folded, then shifted, then multiplied by the other sequence to form the product sequence for that shift, and finally, the values of the product sequence are summed. Except for the folding operation, the computation of the crosscorrelation sequence involves the same operations: shifting one of the sequences, multiplying the two sequences, and summing over all values of the product sequence. Consequently, if we have a computer program that performs convolution, we can use it to perform crosscorrelation by providing as inputs to the program the sequence x(n) and the folded sequence y(−n). Then the convolution of x(n) with y(−n) yields the crosscorrelation rxy (l), that is, rxy (l) = x(l) ∗ y(−l)

(6.8)

We note that the absence of folding makes crosscorrelation a noncommutative operation. In the special case where y(n) = x(n), we have the autocorrelation of x(n), which is defined as the sequence rxx (l) =

∞ 

x(n)x(n − l)

(6.9)

n=−∞

121

Discrete-Time Signals and Systems

or, equivalently, as

∞ 

rxx (l) =

x(n + l)x(n)

(6.10)

n=−∞

In dealing with finite-duration sequences, it is customary to express the autocorrelation and crosscorrelation in terms of the finite limits on the summation. In particular, if x(n) and y(n) are causal sequences of length N [i.e., x(n) = y(n) = 0 for n < 0 and n ≥ N ], the crosscorrelation and autocorrelation sequences may be expressed as N−|k|−1  rxy (l) = (6.11) x(n)y(n − l) n=l

and rxx (l) =

N−|k|−1 

x(n)x(n − l)

(6.12)

n=i

where i = l , k = 0 for l ≥ 0, and i = 0, k = l for l < 0.

6.2

Properties of the Autocorrelation and Crosscorrelation Sequences

The autocorrelation and crosscorrelation sequences have a number of important properties that we now present. To develop these properties, let us assume that we have two sequences x(n) and y(n) with finite energy from which we form the linear combination, ax(n) + by(n − l) where a and b are arbitrary constants and l is some time shift. The energy in this signal is ∞ 

[ax(n) + by(n − l)]2 = a 2

n=−∞

∞ 

x 2 (n) + b2

n=−∞

+ 2ab

∞ 

∞ 

y 2 (n − l)

n=−∞

x(n)y(n − l)

(6.13)

n=−∞

= a rxx (0) + b2 ryy (0) + 2abrxy (l) 2

First, we note that rxx (0) = Ex and ryy (0) = Ey , which are the energies of x(n) and y(n), respectively. It is obvious that a 2 rxx (0) + b2 ryy (0) + 2abrxy (l) ≥ 0 Now, assuming that b = 0, we can divide (6.14) by b 2 to obtain

a 

a  rxx (0) 2 + 2rxy (l) + ryy (0) ≥ 0 b b

122

(6.14)

Discrete-Time Signals and Systems

We view this equation as a quadratic with coefficients rxx (0), 2rxy (l), and ryy (0). Since the quadratic is nonnegative, it follows that the discriminant of this quadratic must be nonpositive, that is, 2 4[rxy (l) − rxx (0)ryy (0)] ≤ 0

Therefore, the crosscorrelation sequence satisfies the condition that   |rxy (l)| ≤ rxx (0)ryy (0) = Ex Ey

(6.15)

In the special case where y(n) = x(n), (6.15) reduces to |rxx (l)| ≤ rxx (0) = Ex

(6.16)

This means that the autocorrelation sequence of a signal attains its maximum value at zero lag. This result is consistent with the notion that a signal matches perfectly with itself at zero shift. In the case of the crosscorrelation sequence, the upper bound on its values is given in (6.15). Note that if any one or both of the signals involved in the crosscorrelation are scaled, the shape of the crosscorrelation sequence does not change; only the amplitudes of the crosscorrelation sequence are scaled accordingly. Since scaling is unimportant, it is often desirable, in practice, to normalize the autocorrelation and crosscorrelation sequences to the range from −1 to 1. In the case of the autocorrelation sequence, we can simply divide by rxx (0). Thus the normalized autocorrelation sequence is defined as rxx (l) ρxx (l) = (6.17) rxx (0) Similarly, we define the normalized crosscorrelation sequence ρxy (l) = 

rxy (l) rxx (0)ryy (0)

(6.18)

Now |ρxx (l)| ≤ 1 and |ρxy (l)| ≤ 1, and hence these sequences are independent of signal scaling. Finally, as we have already demonstrated, the crosscorrelation sequence satisfies the property rxy (l) = ryx (−l) With y(n) = x(n), this relation results in the following important property for the autocorrelation sequence rxx (l) = rxx (−l) (6.19) Hence the autocorrelation function is an even function. Consequently, it suffices to compute rxx (l) for l ≥ 0. EXAMPLE 6.2 Compute the autocorrelation of the signal x(n) = a n u(n), 0 < a < 1 Solution. Since x(n) is an infinite-duration signal, its autocorrelation also has infinite duration. We distinguish two cases.

123

Discrete-Time Signals and Systems

If l ≥ 0, from Fig. 6.2 we observe that rxx (l) =

∞ 

x(n)x(n − l) =

∞ 

n=1

a n a n−l = a −l

n=1

∞  (a 2 )n n=1

Since a < 1, the infinite series converges and we obtain rxx (l) =

1 a |l| , 1 − a2

l≥0

x(n) 1

… … −2 −1 0 1 2 3 (a)



n

x(n − l ) 1

l>0

… 0

n

l (b) x(n − l )

1 l N . For all practical purposes, we can assume that y(n) = 0 for n < 0 and n ≥ M . Now the autocorrelation sequence of y(n), using the normalization factor of 1/M , is ryy (l) =

M−1 1  y(n)y(n − l) M

(6.27)

n=0

If we substitute for y(n) from (6.26) into (6.27) we obtain ryy (l) =

M−1 1  [x(n) + w(n)][x(n − l) + w(n − l)] M n=0

=

M−1 1  x(n)x(n − l) M n=0

+

M−1 1  [x(n)w(n − l) + w(n)x(n − l)] M

(6.28)

n=0

+

M−1 1  w(n)w(n − l) M n=0

= rxx (l) + rxw (l) + rwx (l) + rww (l) The first factor on the right-hand side of (6.28) is the autocorrelation sequence of x(n). Since x(n) is periodic, its autocorrelation sequence exhibits the same periodicity, thus containing relatively large peaks at l = 0, N , 2N , and so on. However, as the shift l approaches M , the peaks are reduced in amplitude due to the fact that we have a finite data record of M samples so that many of the products x(n)x(n − l) are zero. Consequently, we should avoid computing ryy (l) for large lags, say, l > M/2.

126

Discrete-Time Signals and Systems

The crosscorrelations rxw (l) and rwx (l) between the signal x(n) and the additive random interference are expected to be relatively small as a result of the expectation that x(n) and w(n) will be totally unrelated. Finally, the last term on the right-hand side of (6.28) is the autocorrelation sequence of the random sequence w(n). This correlation sequence will certainly contain a peak at l = 0, but because of its random characteristics, rww (l) is expected to decay rapidly toward zero. Consequently, only rxx (l) is expected to have large peaks for l > 0. This behavior allows us to detect the presence of the periodic signal x(n) buried in the interference w(n) and to identify its period. An example that illustrates the use of autocorrelation to identify a hidden periodicity in an observed physical signal is shown in Fig. 6.3. This figure illustrates the autocorrelation (normalized) sequence for the Wölfer sunspot numbers in the 100-year period 1770–1869 for 0 ≤ l ≤ 20, where any value of l corresponds to one year. There is clear evidence in this figure that a periodic trend exists, with a period of 10 to 11 years. EXAMPLE 6.3 Suppose that a signal sequence x(n) = sin(π/5)n, for 0 ≤ n ≤ 99 is corrupted by an additive noise sequence w(n), where the values of the additive noise are selected independently from sample to sample, from a uniform distribution over the range (− /2, /2), where is a parameter of the distribution. The observed sequence is y(n) = x(n) + w(n). Determine the autocorrelation sequence ryy (l) and thus determine the period of the signal x(n). Solution. The assumption is that the signal sequence x(n) has some unknown period that we are attempting to determine from the noise-corrupted observations {y(n)}. Although x(n) is periodic with period 10, we have only a finite-duration sequence of length M = 100 [i.e., 10 periods of x(n)]. The noise power level Pw in the sequence w(n) is determined by the parameter . We simply state that Pw = 2 /12. The signal power level is Px = 21 . Therefore, the signal-to-noise ratio (SNR) is defined as 1 Px 6 = 22 = 2 Pw

/12

Usually, the SNR is expressed on a logarithmic scale in decibels (dB) as 10 log10 (Px /Pw ). Figure 6.4 illustrates a sample of a noise sequence w(n), and the observed sequence y(n) = x(n) + w(n) when the SNR = 1 dB. The autocorrelation sequence ryy (l) is illustrated in Fig. 6.4(c). We observe that the periodic signal x(n), embedded in y(n), results in a periodic autocorrelation function rxx (l) with period N = 10. The effect of the additive noise is to add to the peak value at l = 0, but for l = 0, the correlation sequence rww (l) ≈ 0 as a result of the fact that values of w(n) were generated independently. Such noise is usually called white noise. The presence of this noise explains the reason for the large peak at l = 0. The smaller, nearly equal peaks at l = ±10, ±20, . . . are due to the periodic characteristics of x(n).

6.4

Input–Output Correlation Sequences

In this section we derive two input–output relationships for LTI systems in the “correlation domain.” Let us assume that a signal x(n) with known autocorrelation rxx (l)

127

Discrete-Time Signals and Systems

160 140

Number of Sunspots

120 100 80 60 40 20 0 1770

1790

1810

1830

1850

1870

Year (a) rxx(l) 1 0.8 0.6 0.4 0.2 5 0

15

20

10

l

Lag

Years (b)

Identification of periodicity in the Wölfer sunspot numbers: (a) annual Wölfer sunspot numbers; (b) normalized autocorrelation sequence.

Figure 6.3

is applied to an LTI system with impulse response h(n), producing the output signal y(n) = h(n) ∗ x(n) =

∞ 

h(k)x(n − k)

k=−∞

The crosscorrelation between the output and the input signal is ryx (l) = y(l) ∗ x(−l) = h(l) ∗ [x(l) ∗ x(−l)]

128

Discrete-Time Signals and Systems

ω(n)

n

(a)

y(n)

n

(b)

ryy(l)

SNR = 1 dB

0

l

(c)

Figure 6.4 Use of autocorrelation to detect the presence of a periodic signal

corrupted by noise.

or ryx (l) = h(l) ∗ rxx (l)

(6.29)

where we have used (6.8) and the properties of convolution. Hence the crosscorrelation between the input and the output of the system is the convolution of the impulse response with the autocorrelation of the input sequence. Alternatively, ryx (l) may be viewed as the output of the LTI system when the input sequence is rxx (l). This is illustrated in Fig. 6.5. If we replace l by −l in (6.29), we obtain rxy (l) = h(−l) ∗ rxx (l) The autocorrelation of the output signal can be obtained by using (6.8) with x(n) = y(n) and the properties of convolution. Thus we have ryy (l) = y(l) ∗ y(−l) = [h(l) ∗ x(l)] ∗ [h(−l) ∗ x(−l)] = [h(l) ∗ h(−l)] ∗ [x(l) ∗ x(−l)]

(6.30)

= rhh (l) ∗ rxx (l)

129

Discrete-Time Signals and Systems

The autocorrelation rhh (l) of the impulse response h(n) exists if the system is stable. Furthermore, the stability insures that the system does not change the type (energy or power) of the input signal. By evaluating (6.30) for l = 0 we obtain ryy (0) =

∞ 

(6.31)

rhh (k)rxx (k)

k=−∞

which provides the energy (or power) of the output signal in terms of autocorrelations. These relationships hold for both energy and power signals. The direct derivation of these relationships for energy and power signals, and their extensions to complex signals, are left as exercises for the student.

Input

Figure 6.5

rxx(n)

LTI SYSTEM h(n)

Output ryx(n)

Input–output relation for crosscorrelation ryx (n).

7

Summary and References The major theme of this chapter is the characterization of discrete-time signals and systems in the time domain. Of particular importance is the class of linear timeinvariant (LTI) systems which are widely used in the design and implementation of digital signal processing systems. We characterized LTI systems by their unit sample response h(n) and derived the convolution summation, which is a formula for determining the response y(n) of the system characterized by h(n) to any given input sequence x(n). The class of LTI systems characterized by linear difference equations with constant coefficients is by far the most important of the LTI systems in the theory and application of digital signal processing. The general solution of a linear difference equation with constant coefficients was derived in this chapter and shown to consist of two components: the solution of the homogeneous equation, which represents the natural response of the system when the input is zero, and the particular solution, which represents the response of the system to the input signal. From the difference equation, we also demonstrated how to derive the unit sample response of the LTI system. Linear time-invariant systems were generally subdivided into FIR (finite-duration impulse response) and IIR (infinite-duration impulse response) depending on whether h(n) has finite duration or infinite duration, respectively. The realizations of such systems were briefly described. Furthermore, in the realization of FIR systems, we made the distinction between recursive and nonrecursive realizations. On the other hand, we observed that IIR systems can be implemented recursively, only.

130

Discrete-Time Signals and Systems

There are a number of texts on discrete-time signals and systems. We mention as examples the books by McGillem and Cooper (1984), Oppenheim and Willsky (1983), and Siebert (1986). Linear constant-coefficient difference equations are treated in depth in the books by Hildebrand (1952) and Levy and Lessman (1961). The last topic in this chapter, on correlation of discrete-time signals, plays an important role in digital signal processing, especially in applications dealing with digital communications, radar detection and estimation, sonar, and geophysics. In our treatment of correlation sequences, we avoided the use of statistical concepts. Correlation is simply defined as a mathematical operation between two sequences, which produces another sequence, called either the crosscorrelation sequence when the two sequences are different, or the autocorrelation sequence when the two sequences are identical. In practical applications in which correlation is used, one (or both) of the sequences is (are) contaminated by noise and, perhaps, by other forms of interference. In such a case, the noisy sequence is called a random sequence and is characterized in statistical terms. The corresponding correlation sequence becomes a function of the statistical characteristics of the noise and any other interference. Supplementary reading on probabilistic and statistical concepts dealing with correlation can be found in the books by Davenport (1970), Helstrom (1990), Peebles (1987), and Stark and Woods (1994).

Problems 1 A discrete-time signal x(n) is defined as  x(n) =

1 + n3 , −3 ≤ n ≤ −1 1, 0≤n≤3 0, elsewhere

(a) Determine its values and sketch the signal x(n). (b) Sketch the signals that result if we: 1. First fold x(n) and then delay the resulting signal by four samples. 2. First delay x(n) by four samples and then fold the resulting signal. (c) Sketch the signal x(−n + 4). (d) Compare the results in parts (b) and (c) and derive a rule for obtaining the signal x(−n + k) from x(n). (e) Can you express the signal x(n) in terms of signals δ(n) and u(n)? 2

A discrete-time signal x(n) is shown in Fig. P2. Sketch and label carefully each of the following signals.

131

Discrete-Time Signals and Systems

x(n)

Figure P2

−2

1

1

1

1

−1

0

1

2

1 2

1 2

3

4

n

(a) x(n − 2) (b) x(4 − n) (c) x(n + 2) (d) x(n)u(2 − n) (e) x(n − 1)δ(n − 3) (f) x(n2 ) (g) even part of x(n) (h) odd part of x(n) 3 Show that (a) δ(n) = u(n) − u(n − 1)   (b) u(n) = nk=−∞ δ(k) = ∞ k=0 δ(n − k) 4 Show that any signal can be decomposed into an even and an odd component. Is the decomposition unique? Illustrate your arguments using the signal x(n) = {2, 3, 4, 5, 6} ↑

5 Show that the energy (power) of a real-valued energy (power) signal is equal to the sum of the energies (powers) of its even and odd components. 6 Consider the system y(n) = T [x(n)] = x(n2 ) (a) Determine if the system is time invariant. (b) To clarify the result in part (a) assume that the signal  1, 0 ≤ n ≤ 3 x(n) = 0, elsewhere is applied into the system. (1) Sketch the signal x(n). (2) Determine and sketch the signal y(n) = T [x(n)]. (3) Sketch the signal y2 (n) = y(n − 2). (4) Determine and sketch the signal x2 (n) = x(n − 2). (5) Determine and sketch the signal y2 (n) = T [x2 (n)]. (6) Compare the signals y2 (n) and y(n − 2). What is your conclusion? (c) Repeat part (b) for the system y(n) = x(n) − x(n − 1) Can you use this result to make any statement about the time invariance of this system? Why? (d) Repeat parts (b) and (c) for the system y(n) = T [x(n)] = nx(n) 7 A discrete-time system can be

132

Discrete-Time Signals and Systems

(1) Static or dynamic (2) Linear or nonlinear (3) Time invariant or time varying (4) Causal or noncausal (5) Stable or unstable Examine the following systems with respect to the properties above. (a) y(n) = cos[x(n)]  (b) y(n) = n+1 k=−∞ x(k) (c) y(n) = x(n) cos(ω0 n) (d) y(n) = x(−n + 2) (e) y(n) = Trun[x(n)], where Trun[x(n)] denotes the integer part of x(n), obtained by truncation (f) y(n) = Round[x(n)], where Round[x(n)] denotes the integer part of x(n) obtained by rounding Remark: The systems in parts (e) and (f) are quantizers that perform truncation and rounding, respectively. (g) y(n) = |x(n)| (h) y(n) = x(n)u(n) (i) y(n) = x(n) + nx(n + 1) (j) y(n) = x(2n)  x(n), if x(n) ≥ 0 (k) y(n) = 0, if x(n) < 0 (l) y(n) = x(−n) (m) y(n) = sign[x(n)] (n) The ideal sampling system with input xa (t) and output x(n) = xa (nT ), −∞ < n < ∞ 8 Two discrete-time systems T1 and T2 are connected in cascade to form a new system T as shown in Fig. P8. Prove or disprove the following statements. (a) If T1 and T2 are linear, then T is linear (i.e., the cascade connection of two linear systems is linear). (b) If T1 and T2 are time invariant, then T is time invariant. (c) If T1 and T2 are causal, then T is causal. (d) If T1 and T2 are linear and time invariant, the same holds for T . (e) If T1 and T2 are linear and time invariant, then interchanging their order does not change the system T . (f) As in part (e) except that T1 , T2 are now time varying. (Hint: Use an example.) (g) If T1 and T2 are nonlinear, then T is nonlinear. (h) If T1 and T2 are stable, then T is stable. (i) Show by an example that the inverses of parts (c) and (h) do not hold in general.

133

Discrete-Time Signals and Systems

x(n)

T1

T2

y(n)

T = T 1T 2

Figure P8

9 Let T be an LTI, relaxed, and BIBO stable system with input x(n) and output y(n). Show that: (a) If x(n) is periodic with period N [i.e., x(n) = x(n + N ) for all n ≥ 0], the output y(n) tends to a periodic signal with the same period. (b) If x(n) is bounded and tends to a constant, the output will also tend to a constant. (c) If x(n) is an energy signal, the output y(n) will also be an energy signal. 10 The following input–output pairs have been observed during the operation of a timeinvariant system: T

x1 (n) = {1, 0, 2} ←→ y1 (n) = {0, 1, 2} ↑



T

x2 (n) = {0, 0, 3} ←→ y2 (n) = {0, 1, 0, 2} ↑



T

x3 (n) = {0, 0, 0, 1} ←→ y3 (n) = {1, 2, 1} ↑



11

Can you draw any conclusions regarding the linearity of the system. What is the impulse response of the system? The following input–output pairs have been observed during the operation of a linear system: T

x1 (n) = {−1, 2, 1} ←→ y1 (n) = {1, 2, −1, 0, 1} ↑



T

x2 (n) = {1, −1, −1} ←→ y2 (n) = {−1, 1, 0, 2} ↑



T

x3 (n) = {0, 1, 1} ←→ y3 (n) = {1, 2, 1} ↑



Can you draw any conclusions about the time invariance of this system? 12 The only available information about a system consists of N input–output pairs, of signals yi (n) = T [xi (n)], i = 1, 2, . . . , N . (a) What is the class of input signals for which we can determine the output, using the information above, if the system is known to be linear? (b) The same as above, if the system is known to be time invariant. 13 Show that the necessary and sufficient condition for a relaxed LTI system to be BIBO stable is ∞  |h(n)| ≤ Mh < ∞ n=−∞

for some constant Mn .

134

Discrete-Time Signals and Systems

14

Show that: (a) A relaxed linear system is causal if and only if for any input x(n) such that x(n) = 0 for n < n0 ⇒ y(n) = 0

for n < n0

(b) A relaxed LTI system is causal if and only if h(n) = 0

for n < 0

15 (a) Show that for any real or complex constant a , and any finite integer numbers M and N , we have N 

 n = Ma = n

a M − a N+1 , if a = 1 1−a N − M + 1, if a = 1

(b) Show that if |a| < 1, then ∞ 

an =

n=0

16

(a) If y(n) = x(n) ∗ h(n), show that

 y

=

1 1−a  

h,

x

where

 x

=

∞ n=−∞

x(n).

(b) Compute the convolution y(n) = x(n) ∗ h(n) of the following signals and check the correctness of the results by using the test in (a). (1) x(n) = {1, 2, 4}, h(n) = {1, 1, 1, 1, 1} (2) x(n) = {1, 2, −1}, h(n) = x(n) (3) x(n) = {0, 1, −2, 3, −4}, h(n) = { 21 , 21 , 1, 21 } (4) x(n) = {1, 2, 3, 4, 5}, h(n) = {1} (5) x(n) = {1, −2, 3}, h(n) = {0, 0, 1, 1, 1, 1} ↑



(6) x(n) = {0, 0, 1, 1, 1, 1}, h(n) = {1, −2, 3} ↑



(7) x(n) = {0, 1, 4, −3}, h(n) = {1, 0, −1, −1} ↑



(8) x(n) = {1, 1, 2}, h(n) = u(n) ↑

(9) x(n) = {1, 1, 0, 1, 1}, h(n) = {1, −2, −3, 4} ↑



(10) x(n) = {1, 2, 0, 2, 1}h(n) = x(n) ↑

(11) x(n) =

( 21 )n u(n), h(n)

= ( 41 )n u(n)

135

Discrete-Time Signals and Systems

17 Compute and plot the convolutions x(n) ∗ h(n) and h(n) ∗ x(n) for the pairs of signals shown in Fig. P17. x(n)

h(n) 6

1 0123

n

01234 5 6 (a)

x(n)

n

h(n) 6

1 0123

−3 −2 −1 0 1 2 3

n (b)

x(n)

h(n)

1

1 3 4 56

−4 −3

n

n

(c)

x(n)

h(n)

1

Figure P17

1 23 4 5

−2 −1

n (d)

18 Determine and sketch the convolution y(n) of the signals  1 x(n) = 3 n, 0 ≤ n ≤ 6 0, elsewhere  1, −2 ≤ n ≤ 2 h(n) = 0, elsewhere (a) Graphically (b) Analytically 19 Compute the convolution y(n) of the signals  n α , −3 ≤ n ≤ 5 x(n) = 0, elsewhere  1, 0 ≤ n ≤ 4 h(n) = 0, elsewhere 20

136

n

Consider the following three operations. (a) Multiply the integer numbers: 131 and 122. (b) Compute the convolution of signals: {1, 3, 1} ∗ {1, 2, 2}. (c) Multiply the polynomials: 1 + 3z + z2 and 1 + 2z + 2z2 . (d) Repeat part (a) for the numbers 1.31 and 12.2. (e) Comment on your results.

n

Discrete-Time Signals and Systems

21

Compute the convolution y(n) = x(n) ∗ h(n) of the following pairs of signals. (a) x(n) = a n u(n), h(n) = bn u(n) when a = b and when a = b  1, n = −2, 0, 1 (b) x(n) = 2, n = −1 h(n) = δ(n) − δ(n − 1) + δ(n − 4) + δ(n − 5) 0, elsewhere (c) x(n) = u(n + 1) − u(n − 4) − δ(n − 5); h(n) = [u(n + 2) − u(n − 3)] · (3 − |n|) (d) x(n) = u(n) − u(n − 5); h(n) = u(n − 2) − u(n − 8) + u(n − 11) − u(n − 17)

22 Let x(n) be the input signal to a discrete-time filter with impulse response hi (n) and let yi (n) be the corresponding output. (a) Compute and sketch x(n) and yi (n) in the following cases, using the same scale in all figures. x(n) = {1, 4, 2, 3, 5, 3, 3, 4, 5, 7, 6, 9} h1 (n) = {1, 1} h2 (n) = {1, 2, 1} 1 1 h3 (n) = { , } 2 2 1 1 1 h4 (n) = { , , } 4 2 4 1 1 1 h5 (n) = { , − , } 4 2 4 Sketch x(n), y1 (n), y2 (n) on one graph and x(n), y3 (n), y4 (n), y5 (n) on another graph (b) What is the difference between y1 (n) and y2 (n), and between y3 (n) and y4 (n)? (c) Comment on the smoothness of y2 (n) and y4 (n). Which factors affect the smoothness? (d) Compare y4 (n) with y5 (n). What is the difference? Can you explain it? (e) Let h6 (n) = { 21 , − 21 }. Compute y6 (n). Sketch x(n), y2 (n), and y6 (n) on the same figure and comment on the results. 23

Express the output y(n) of a linear time-invariant system with impulse response h(n) in terms of its step response s(n) = h(n)∗ u(n) and the input x(n). 24 The discrete-time system y(n) = ny(n − 1) + x(n),

n≥0

is at rest [i.e., y(−1) = 0]. Check if the system is linear time invariant and BIBO stable. 25 Consider the signal γ (n) = a n u(n), 0 < a < 1.

137

Discrete-Time Signals and Systems

(a) Show that any sequence x(n) can be decomposed as x(n) =

∞ 

ck γ (n − k)

n=−∞

and express ck in terms of x(n). (b) Use the properties of linearity and time invariance to express the output y(n) = T [x(n)] in terms of the input x(n) and the signal g(n) = T [γ (n)], where T [·] is an LTI system. (c) Express the impulse response h(n) = T [δ(n)] in terms of g(n). 26 Determine the zero-input response of the system described by the second-order difference equation x(n) − 3y(n − 1) − 4y(n − 2) = 0 27 Determine the particular solution of the difference equation 5 1 y(n − 1) − y(n − 2) + x(n) 6 6 n when the forcing function is x(n) = 2 u(n). 28 In Example 4.8, equation (4.30), separate the output s equence y(n) into the transient response and the steady-state response. Plot these two responses for a1 = −0.9. 29 Determine the impulse response for the cascade of two linear time-invariant systems having impulse responses. y(n) =

h1 (n) = a n [u(n) − u(n − N )] and h2 (n) = [u(n) − u(n − M)] 30 Determine the response y(n), n ≥ 0, of the system described by the second-order difference equation y(n) − 3y(n − 1) − 4y(n − 2) = x(n) + 2x(n − 1) to the input x(n) = 4n u(n). 31 Determine the impulse response of the following causal system: y(n) − 3y(n − 1) − 4y(n − 2) = x(n) + 2x(n − 1) 32 Let x(n), N1 ≤ n ≤ N2 and h(n), M1 ≤ n ≤ M2 be two finite-duration signals. (a) Determine the range L1 ≤ n ≤ L2 of their convolution, in terms of N1 , N2 , M1 and M2 . (b) Determine the limits of the cases of partial overlap from the left, full overlap, and partial overlap from the right. For convenience, assume that h(n) has shorter duration than x(n). (c) Illustrate the validity of your results by computing the convolution of the signals  1, −2 ≤ n ≤ 4 x(n) = 0, elsewhere  2, −1 ≤ n ≤ 2 h(n) = 0, elsewhere

138

Discrete-Time Signals and Systems

33

Determine the impulse response and the unit step response of the systems described by the difference equation (a) y(n) = 0.6y(n − 1) − 0.08y(n − 2) + x(n) (b) y(n) = 0.7y(n − 1) − 0.1y(n − 2) + 2x(n) − x(n − 2)

34

Consider a system with impulse response  h(n) =

( 21 )n , 0 ≤ n ≤ 4 0, elsewhere

Determine the input x(n) for 0 ≤ n ≤ 8 that will generate the output sequence y(n) = {1, 2, 2.5, 3, 3, 3, 2, 1, 0, . . .} ↑

35 Consider the interconnection of LTI systems as shown in Fig. P35. h2(n) y(n)

x(n) +

h1(n) ← h3(n)

h4(n)

Figure P35

(a) Express the overall impulse response in terms of h1 (n), h2 (n), h3 (n), and h4 (n). (b) Determine h(n) when 1 1 1 h1 (n) = { , , } 2 4 2 h2 (n) = h3 (n) = (n + 1)u(n) h4 (n) = δ(n − 2) (c) Determine the response of the system in part (b) if x(n) = δ(n + 2) + 3δ(n − 1) − 4δ(n − 3) 36 Consider the system in Fig. P36 with h(n) = a n u(n), −1 < a < 1. Determine the response y(n) of the system to the excitation x(n) = u(n + 5) − u(n − 10)

139

Discrete-Time Signals and Systems

h(n) x(n)

y(n) + − z−2

h(n)

Figure P36

37 Compute and sketch the step response of the system y(n) =

M−1 1  x(n − k) M k=0

38 Determine the range of values of the parameter a for which the linear time-invariant system with impulse response  h(n) =

a n , n ≥ 0, n even 0, otherwise

is stable. 39 Determine the response of the system with impulse response h(n) = a n u(n) to the input signal x(n) = u(n) − u(n − 10)

40

(Hint: The solution can be obtained easily and quickly by applying the linearity and time-invariance properties to the result in Example 3.5.) Determine the response of the (relaxed) system characterized by the impulse response 1 h(n) = ( )n u(n) 2 to the input signal

 x(n) =

1, 0 ≤ n < 10 0, otherwise

41 Determine the response of the (relaxed) system characterized by the impulse response 1 h(n) = ( )n u(n) 2 to the input signals (a) x(n) = 2n u(n) (b) x(n) = u(−n)

140

Discrete-Time Signals and Systems

42

Three systems with impulse responses h1 (n) = δ(n) − δ(n − 1), h2 (n) = h(n), and h3 (n) = u(n), are connected in cascade. (a) What is the impulse response, hc (n), of the overall system? (b) Does the order of the interconnection affect the overall system?

43

(a) Prove and explain graphically the difference between the relations x(n)δ(n − n0 ) = x(n0 )δ(n − n0 )

and

x(n) ∗ δ(n − n0 ) = x(n − n0 )

(b) Show that a discrete-time system, which is described by a convolution summation, is LTI and relaxed, (c) What is the impulse response of the system described by y(n) = x(n − n0 )? 44

Two signals s(n) and v(n) are related through the following difference equations. s(n) + a1 s(n − 1) + · · · + aN s(n − N ) = b0 v(n) Design the block diagram realization of: (a) The system that generates s(n) when excited by v(n). (b) The system that generates v(n) when excited by s(n). (c) What is the impulse response of the cascade interconnection of systems in parts (a) and (b)?

45

Compute the zero-state response of the system described by the difference equation 1 y(n) + y(n − 1) = x(n) + 2x(n − 2) 2 to the input x(n) = {1, 2, 3, 4, 2, 1} ↑

by solving the difference equation recursively. 46 Determine the direct form II realization for each of the following LTI systems: (a) 2y(n) + y(n − 1) − 4y(n − 3) = x(n) + 3x(n − 5) (b) y(n) = x(n) − x(n − 1) + 2x(n − 2) − 3x(n − 4) 47 Consider the discrete-time system shown in Fig. P47. x(n)

+

+

y(n)

z−1

Figure P47

1 2

141

Discrete-Time Signals and Systems

(a) Compute the 10 first samples of its impulse response. (b) Find the input–output relation. (c) Apply the input x(n) = {1, 1, 1, . . .} and compute the first 10 samples of the ↑

output. (d) Compute the first 10 samples of the output for the input given in part (c) by using convolution. (e) Is the system causal? Is it stable? 48 Consider the system described by the difference equation y(n) = ay(n − 1) + bx(n) (a) Determine b in terms of a so that ∞ 

h(n) = 1

n=−∞

(b) Compute the zero-state step response s(n) of the system and choose b so that s(∞) = 1. (c) Compare the values of b obtained in parts (a) and (b). What did you notice? 49 A discrete-time system is realized by the structure shown in Fig. P49. (a) Determine the impulse response. (b) Determine a realization for its inverse system, that is, the system which produces x(n) as an output when y(n) is used as an input. x(n)

2

+

+

y(n)

+

y(n)

z−1 3

Figure P49 0.8

50

Consider the discrete-time system shown in Fig. P50. x(n) + z−1 0.9

2 + z−1

Figure P50

142

3

Discrete-Time Signals and Systems

(a) Compute the first six values of the impulse response of the system. (b) Compute the first six values of the zero-state step response of the system. (c) Determine an analytical expression for the impulse response of the system. 51

Determine and sketch the impulse response of the following systems for n = 0, 1, . . . , 9. x(n)

z−1

z−1

z−1

+

+

y(n)

1 3 z−1

(a) x(n)

z−1

1 2

z−1

+

y(n) z−1

1 2

+

z−1

1 8 (b) x(n) +

+

y(n)

z−1

z−1

0.8

0.6 (c)

Figure P51

(a) Fig. P51(a). (b) Fig. P51(b). (c) Fig. P51(c). (d) Classify the systems above as FIR or IIR. (e) Find an explicit expression for the impulse response of the system in part (c).

143

Discrete-Time Signals and Systems

52

Consider the systems shown in Fig. P52. x(n)

z−1

z−1

c0

c1

c2 +

+

y(n)

x(n) b0

b1 z−1

x(n)

b2 z−1

+

+

z−1

y(n)

z−1 a1

a0

+

y(n)

a2

+

Figure P52

(a) Determine and sketch their impulse responses h1 (n), h2 (n), and h3 (n). (b) Is it possible to choose the coefficients of these systems in such a way that h1 (n) = h2 (n) = h3 (n) 53 Consider the system shown in Fig. P53. x(n)

z−1

+

+

y(n) 1 2 z−1

Figure P53

(a) Determine its impulse response h(n). (b) Show that h(n) is equal to the convolution of the following signals: h1 (n) = δ(n) + δ(n − 1)

54

1 h2 (n) = ( )n u(n) 2 Compute and sketch the convolution yi (n) and correlation ri (n) sequences for the following pair of signals and comment on the results obtained. h1 (n) = {1, 1, 1, 1, 1} (a) x1 (n) = {1, 2, 4} ↑

(b) x2 (n) = {0, 1, −2, 3, −4} ↑

144



h2 (n) = { 21 , 1, 2, 1, 21 } ↑

Discrete-Time Signals and Systems

(c) x3 (n) = {1, 2, 3, 4}

h3 (n) = {4, 3, 2, 1}

(d) x4 (n) = {1, 2, 3, 4}

h4 (n) = {1, 2, 3, 4}









55 The zero-state response of a causal LTI system to the input x(n) = {1, 3, 3, 1} is y(n) = {1, 4, 6, 4, 1}. Determine its impulse response.





56 Prove by direct substitution the equivalence of equations (5.9) and (5.10), which describe the direct form II structure, to the relation (5.6), which describes the direct form I structure. 57 Determine the response y(n), n ≥ 0 of the system described by the second-order difference equation y(n) − 4y(n − 1) + 4y(n − 2) = x(n) − x(n − 1) when the input is x(n) = (−1)n u(n) and the initial conditions are y(−1) = y(−2) = 0. 58 Determine the impulse response h(n) for the system described by the second-order difference equation y(n) − 4y(n − 1) + 4y(n − 2) = x(n) − x(n − 1) 59 Show that any discrete-time signal x(n) can be expressed as x(n) =

∞ 

[x(k) − x(k − 1)]u(n − k)

k=−∞

where u(n − k) is a unit step delayed by k units in time, that is,  1, n ≥ k u(n − k) = 0, otherwise 60 Show that the output of an LTI system can be expressed in terms of its unit step response s(n) as follows. y(n) =

∞ 

[s(k) − s(k − 1)]x(n − k)

k=−∞

=

∞ 

[x(k) − x(k − 1)]s(n − k)

k=−∞

61 Compute the correlation sequences quences.  1, x(n) = 0,  1, y(n) = 0,

rxx (l) and rxy (l) for the following signal sen0 − N ≤ n ≤ n0 + N otherwise −N ≤ n ≤ N otherwise

145

Discrete-Time Signals and Systems

62 Determine the autocorrelation sequences of the following signals. (a) x(n) = {1, 2, 1, 1} ↑

(b) y(n) = {1, 1, 2, 1} ↑

What is your conclusion? 63 What is the normalized autocorrelation sequence of the signal x(n) given by  x(n) =

1, −N ≤ n ≤ N 0, otherwise

64 An audio signal s(t) generated by a loudspeaker is reflected at two different walls with reflection coefficients r1 and r2 . The signal x(t) recorded by a microphone close to the loudspeaker, after sampling, is x(n) = s(n) + r1 s(n − k1 ) + r2 s(n − k2 ) where k1 and k2 are the delays of the two echoes. (a) Determine the autocorrelation rxx (l) of the signal x(n). (b) Can we obtain r1 , r2 , k1 , and k2 by observing rxx (l)? (c) What happens if r2 = 0? 65

Time-delay estimation in radar Let xa (t) be the transmitted signal and ya (t) be the received signal in a radar system, where ya (t) = axa (t − td ) + va (t) and va (t) is additive random noise. The signals xa (t) and ya (t) are sampled in the receiver, according to the sampling theorem, and are processed digitally to determine the time delay and hence the distance of the object. The resulting discrete-time signals are x(n) = xa (nT ) y(n) = ya (nT ) = axa (nT − DT ) + va (nT )

= ax(n − D) + v(n)

1

0

0

0

Output 0 → ⫺1 1 → ⫹1

Figure P65

Linear feedback shift register.

146

+

Modulo-2 adder

Discrete-Time Signals and Systems

(a) Explain how we can measure the delay D by computing the crosscorrelation rxy (l). (b) Let x(n) be the 13-point Barker sequence x(n) = {+1, +1, +1, +1, +1, −1, −1, +1, +1, −1, +1, −1, +1} and v(n) be a Gaussian random sequence with zero mean and variance σ 2 = 0.01. Write a program that generates the sequence y(n), 0 ≤ n ≤ 199 for a = 0.9 and D = 20. Plot the signals x(n), y(n), 0 ≤ n ≤ 199. (c) Compute and plot the crosscorrelation rxy (l), 0 ≤ l ≤ 59. Use the plot to estimate the value of the delay D. (d) Repeat parts (b) and (c) for σ 2 = 0.1 and σ 2 = 1. (e) Repeat parts (b) and (c) for the signal sequence x(n) = {−1, −1, −1, +1, +1, +1, +1, −1, +1, −1, +1, +1, −1, −1, +1} which is obtained from the four-stage feedback shift register shown in Fig. P6.5. Note that x(n) is just one period of the periodic sequence obtained from the feedback shift register. (f) Repeat parts (b) and (c) for a sequence of period N = 27 − 1, which is obtained from a seven-stage feedback shift register. Table 2 gives the stages connected to the modulo-2 adder for (maximal-length) shift-register sequences of length N = 2m − 1. TABLE 2 Shift-Register Connections for Generating Maximal-Length Sequences m Stages Connected to Modulo-2 Adder 1 1 2 1, 2 3 1, 3 4 1, 4 5 1, 4 6 1, 6 7 1, 7 8 1, 5, 6, 7 9 1, 6 10 1, 8 11 1, 10 12 1, 7, 9, 12 13 1, 10, 11, 13 14 1, 5, 9, 14 15 1, 15 16 1, 5, 14, 16 17 1, 15

147

Discrete-Time Signals and Systems

66

Implementation of LTI systems Consider the recursive discrete-time system described by the difference equation y(n) = −a1 y(n − 1) − a2 y(n − 2) + b0 x(n) where a1 = −0.8, a2 = 0.64, and b0 = 0.866. (a) Write a program to compute and plot the impulse response h(n) of the system for 0 ≤ n ≤ 49. (b) Write a program to compute and plot the zero-state step response s(n) of the system for 0 ≤ n ≤ 100. (c) Define an FIR system with impulse response hFIR (n) given by  hFIR (n) =

h(n), 0 ≤ n ≤ 19 0, elsewhere

where h(n) is the impulse response computed in part (a). Write a program to compute and plot its step response. (d) Compare the results obtained in parts (b) and (c) and explain their similarities and differences. 67 Write a computer program that computes the overall impulse response h(n) of the system shown in Fig. P67 for 0 ≤ n ≤ 99. The systems T1 , T2 , T3 , and T4 are specified by 1 1 1 1 1 T1 : h1 (n) = {1, , , , , } ↑ 2 4 8 16 32 T2 : h2 (n) = {1, 1, 1, 1, 1} ↑

T3 : y3 (n) =

1 1 1 x(n) + x(n − 1) + x(n − 2) 4 2 4

T4 : y(n) = 0.9y(n − 1) − 0.81y(n − 2) + v(n) + v(n − 1) Plot h(n) for 0 ≤ n ≤ 99. T1

T2

x(n) + T3

Figure P67

148

y3(n)

ν(n)

T4

y(n)

Discrete-Time Signals and Systems

Answers to Selected Problems 7

(a) Static, nonlinear, time invariant, causal, stable. (c) Static, linear, time variant, causal, stable. (e) Static, nonlinear, time invariant, causal, stable. (h) Static, linear, time invariant, causal, stable. (k) Static, nonlinear, time invariant, causal, stable.

11

Since the system is linear and x1 (n) + x2 (n) = δ(n), it follows that the impulse response of the system is y1 (n) + y2 (n) = 0, 3, −1, 2, 1 . ↑   If the system were time invariant, the response to x3 (n) would be 3, 2, 1, 3, 1 . But this is not ↑

the case.

16

19

22 26

32

  = 35, = 5, (b)  (1) y(n) = h(n) ∗ x(n) = {1, 3, 7, 7, 7, 6, 4}; n y(n) k h(k) k x(k) = 7.    1, n x(n) = 15 (4) y(n) = {1, 2, 3, 4, 5}; n y(n) =15, n h(n) =  (7) y(n) = {0, 1, 4, −4, −5, −1, 3}  n y(n) = −2,  n h(n) = −1,  n x(n) = 2 (10) y(n) = {1, 4, 4, 4, 10, 4, 4, 4, 1}; n y(n) = 36, n h(n) = 6, n x(n) = 6      y(n) = 4k=0 h(k)x(n − k), x(n) = a −3 , a −2 , a −1 , 1, a, . . . , a 5 , h(n) = 1, 1, 1, 1, 1, ; y(n) = ↑ ↑ 4 k=0 x(n − k), −3 ≤ n ≤ 9; y(n) = 0, otherwise (a) y1 (n) = x(n) + x(n − 1) = {1, 5, 6, 5, 8, 8, 6, 7, 9, 12, 15, 9} y4 (n) = {0.25, 1.5, 2.75, 2.75, 3.25, 4, 3.5, 3.25, 3.75, 5.25, 6.25, 7, 6, 2.25}

 2 With x(n) = 0, we have y(n − 1) + 43 y(n − 1) = 0 y(−1) = − 43 y(−2); y(0) = − 43 y(−2);  4 3 y(1) = − 3 y(−2)  k+2 y(−2) ← zero-input response. Therefore, y(k) = − 43

(a) L1 = N1 + M1 and L2 = N2 + M2 (c) N1 = −2, N2 = 4, M1 = −1, M2 = 2

34

Partial overlap from left: n = −3 n = −1 L1 = −3; Full overlap: n = 0 n = 3; Partial overlap from right: n = 4 n = 6 L2 = 6     h(n) = 1, 21 , 41 , 18 , 161 ; y(n) = 1, 2, 2, 5, 3, 3, 3, 2, 1, 0

38

 Then, x(n) = 1. 23 , 23 , 47 , 23 ∞ ∞ ∞ n 2n n=−∞ |h(n)| = n=0,neven |a| ; n=− |a| ; =

40 41 42





1 1−|a |2

Stable if |a| < 1



y(n) = 2 1 − ( 21 )n+1 u(n) − 2 1 − ( 21 )a−9 u(n − 10)

(a) y(n) = 23 2n+1 − ( 21 )n+1 u(n) (a) hc (n) = h1 (n) ∗ h2 (n) ∗ h3 (n) = [δ(n) − δ(n − 1)] ∗ u(n) ∗ h(n) = h(n) (b) No.

45 48

, y(1) = 47 ,··· y(n) = − 21 y(n − 1) + z(n) + 2x(n − 2); y(−2) = 1, y(−1) = 23 , y(9) = 17 4 8  ∞ b (a) y(n) = ay(n − 1) + bx(n) ⇒ h(n) = ba n u(n) n=0 h(n) = 1−a = 1 ⇒ b = 1 − a. n+1  (b) s(n) = nk=0 h(n − k) = b 1−a u(n) 1−a s(∞) =

51

b 1−a

=1⇒b =1−a

(a) y(n) = 31 x(n) + 13 x(n − 3) + y(n − 1) for x(n) = δ(n), we have  h(n) = 13 , 13 , 13 , 23 , 23 , 23 , 23 . . . (b) y(n) = 21 y(n − 1) + 18 y(n − 2) + 21 x(n − 2), y(−1) = y(−2) = 0  11 15 41 with x(n) = δ(n), h(n) = 0, 0, 21 , 41 , 163 , 19 , 128 , 256 , 1024 ... (c) y(n) = 1.4y(n − 1) − 0.48y(n − 2) + x(n), y(−1) − y(−2) = 0 with x(n)δ(n), h(n) = {1, 1, 4, 1.48, 1.4, 1.2496, 1.0774, 0.9086, . . . } (d) All three systems are IIR.

149

Discrete-Time Signals and Systems

54

  (a) convolution: y1 (n) = 1 3, 7, 7, 7, 7, 4 →   correlation: γ1 (n) = 1, 3, 7, 7, 7 , 6, 4 →   (c) convolution: y4 (n) = 1 , 4, 10, 20, 25, 24, 16 →   correlation: γ4 (n) = 4, 11, 20, 30, 20, 11, 4 →

58

61

63

150

h(n) = [c1 2n + c2 n2n ]u(n) With y(0) = 1, y(1) = 3, we have, c1 = 1, and c2 = 21 .  1, n0 − N ≤ n ≤ n0 + N x(n) = . 0 otherwise  2N + 1 − |l|, −2N ≤ l ≤ 2N rxx (l) = . 0, otherwise   2N + 1 − |l|, −2N ≤ l ≤ 2N . rxx (l) = ∞ n=−∞ x(n)x(n − l) = 0, otherwise Since rxx  (0) = 2N + 1, the normalized autocorrelation is 1 (2N + 1 − |l|), −2N ≤ l ≤ 2N ρxx (l) = 2N+1 0, otherwise

The z -Transform and Its Application to the Analysis of LTI Systems

Transform techniques are an important tool in the analysis of signals and linear timeinvariant (LTI) systems. In this chapter we introduce the z-transform, develop its properties, and demonstrate its importance in the analysis and characterization of linear time-invariant systems. The z-transform plays the same role in the analysis of discrete-time signals and LTI systems as the Laplace transform does in the analysis of continuous-time signals and LTI systems. For example, we shall see that in the z-domain (complex z-plane) the convolution of two time-domain signals is equivalent to multiplication of their corresponding z-transforms. This property greatly simplifies the analysis of the response of an LTI system to various signals. In addition, the z-transform provides us with a means of characterizing an LTI system, and its response to various signals, by its pole–zero locations. We begin this chapter by defining the z-transform. Its important properties are presented in Section 2. In Section 3 the transform is used to characterize signals in terms of their pole–zero patterns. Section 4 describes methods for inverting the z-transform of a signal so as to obtain the time-domain representation of the signal. Finally, in Section 6, we treat one-sided z-transform and use it to solve linear difference equations with nonzero initial conditions. Section 5 is focused on the use of the z-transform in the analysis of LTI systems.

1

The z -Transform In this section we introduce the z-transform of a discrete-time signal, investigate its convergence properties, and briefly discuss the inverse z-transform.

From Chapter 3 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

151

The z -Transform and Its Application to the Analysis of LTI Systems

1.1

The Direct z -Transform

The z-transform of a discrete-time signal x(n) is defined as the power series X(z) ≡

∞ 

x(n)z−n

(1.1)

n=−∞

where z is a complex variable. The relation (1.1) is sometimes called the direct z transform because it transforms the time-domain signal x(n) into its complex-plane representation X(z). The inverse procedure [i.e., obtaining x(n) from X(z)] is called the inverse z -transform and is examined briefly in Section 1.2 and in more detail in Section 4. For convenience, the z-transform of a signal x(n) is denoted by X(z) ≡ Z{x(n)}

(1.2)

whereas the relationship between x(n) and X(z) is indicated by z

x(n) ←→ X(z)

(1.3)

Since the z-transform is an infinite power series, it exists only for those values of z for which this series converges. The region of convergence (ROC) of X(z) is the set of all values of z for which X(z) attains a finite value. Thus any time we cite a z-transform we should also indicate its ROC. We illustrate these concepts by some simple examples. EXAMPLE

1.1

Determine the z-transforms of the following finite-duration signals. (a) x1 (n) = {1, 2, 5, 7, 0, 1} ↑

(b) x2 (n) = {1, 2, 5, 7, 0, 1} ↑

(c) x3 (n) = {0, 0, 1, 2, 5, 7, 0, 1} ↑

(d) x4 (n) = {2, 4, 5, 7, 0, 1} ↑

(e) x5 (n) = δ(n) (f) x6 (n) = δ(n − k), k > 0 (g) x7 (n) = δ(n + k), k > 0 Solution. From definition (1.1), we have (a) X1 (z) = 1 + 2z−1 + 5z−2 + 7z−3 + z−5 , ROC: entire z-plane except z = 0 (b) X2 (z) = z2 + 2z + 5 + 7z−1 + z−3 , ROC: entire z-plane except z = 0 and z = ∞ (c) X3 (z) = z−2 + 2z−3 + 5z−4 + 7z−5 + z−7 , ROC: entire z-plane except z = 0 (d) X4 (z) = 2z2 + 4z + 5 + 7z−1 + z−3 , ROC: entire z-plane except z = 0 and z = ∞ z

(e) X5 (z) = 1 [i.e., δ(n) ←→ 1], ROC: entire z-plane z

(f) X6 (z) = z−k [i.e., δ(n − k) ←→ z−k ], k > 0, ROC: entire z-plane except z = 0 z

(g) X7 (z) = zk [i.e., δ(n + k) ←→ zk ], k > 0, ROC: entire z-plane except z = ∞

152

The z -Transform and Its Application to the Analysis of LTI Systems

From this example it is easily seen that the ROC of a finite-duration signal is the entire z-plane, except possibly the points z = 0 and/or z = ∞. These points are excluded, because zk (k > 0) becomes unbounded for z = ∞ and z−k (k > 0) becomes unbounded for z = 0. From a mathematical point of view the z-transform is simply an alternative representation of a signal. This is nicely illustrated in Example 1.1, where we see that the coefficient of z−n , in a given transform, is the value of the signal at time n. In other words, the exponent of z contains the time information we need to identify the samples of the signal. In many cases we can express the sum of the finite or infinite series for the ztransform in a closed-form expression. In such cases the z-transform offers a compact alternative representation of the signal. EXAMPLE 1.2 Determine the z-transform of the signal 1 x(n) = ( )n u(n) 2 Solution.

The signal x(n) consists of an infinite number of nonzero values 1 1 1 1 x(n) = {1, ( ), ( )2 , ( )3 , . . . , ( )n , . . .} 2 2 2 2

The z-transform of x(n) is the infinite power series 1 1 1 X(z) = 1 + z−1 + ( )2 z−2 + ( )n z−n + · · · 2 2 2 =

∞ ∞   1 1 ( )n z−n = ( z−1 )n 2 2 n=0

n=0

This is an infinite geometric series. We recall that 1 + A + A 2 + A3 + · · · =

1 1−A

if |A| < 1

Consequently, for | 21 z−1 | < 1, or equivalently, for |z| > 21 , X(z) converges to X(z) =

1 , 1 − 1 z−1 2

ROC: |z| >

1 2

We see that in this case, the z-transform provides a compact alternative representation of the signal x(n).

153

The z -Transform and Its Application to the Analysis of LTI Systems

Let us express the complex variable z in polar form as z = rej θ

(1.4)

where r = |z| and θ = ⭿z. Then X(z) can be expressed as X(z)|z=rej θ =

∞ 

x(n)r −n e−j θn

n=−∞

In the ROC of X(z), |X(z)| < ∞. But   ∞      x(n)r −n e−j θn  |X(z)| =   n=−∞ ≤

∞ 

|x(n)r −n e−j θn | =

n=−∞

∞ 

(1.5) |x(n)r −n |

n=−∞

Hence |X(z)| is finite if the sequence x(n)r −n is absolutely summable. The problem of finding the ROC for X(z) is equivalent to determining the range of values of r for which the sequence x(n)r −n is absolutely summable. To elaborate, let us express (1.5) as |X(z)| ≤

−1  n=−∞



∞  n=1

|x(n)r −n | +

 ∞    x(n)     rn  n=0

 ∞    x(n)    |x(−n)r n | +  rn 

(1.6)

n=0

If X( z) converges in some region of the complex plane, both summations in (1.6) must be finite in that region. If the first sum in (1.6) converges, there must exist values of r small enough such that the product sequence x(−n)r n , 1 ≤ n < ∞, is absolutely summable. Therefore, the ROC for the first sum consists of all points in a circle of some radius r1 , where r1 < ∞, as illustrated in Fig. 1.1(a). On the other hand , if the second sum in (1.6) converges, there must exist values of r large enough such that the product sequence x(n)/r n , 0 ≤ n < ∞, is absolutely summable. Hence the ROC for the second sum in (1.6) consists of all points outside a circle of radius r > r2 , as illustrated in Fig. 1.1(b). Since the convergence of X( z) requires that both sums in (1.6) be finite, it follows that the ROC of X(z) is generally specified as the annular region in the zplane, r2 < r < r1 , which is the common region where both sums are finite. This region is illustrated in Fig. 1.1(c). On the other hand, if r2 > r1 , there is no common region of convergence for the two sums and hence X(z) does not exist. The following examples illustrate these important concepts.

154

The z -Transform and Its Application to the Analysis of LTI Systems

Im(z)

z-plane

r1

Re(z)

Region of convergence for

Σ

|x(−n) r n |

n=1

(a) Im(z)

z-plane

r2

Re(z)

Region of convergence for x(n) rn n=0

Σ

(b) Im(z)

z-plane

r2

Re(z)

r1

Figure 1.1

Region of convergence for X(z) and its corresponding causal and anticausal components.

Region of convergence for |X(z)| r2 < r < r1 (c)

EXAMPLE 1.3 Determine the z-transform of the signal  x(n) = α n u(n) = Solution.

αn , n ≥ 0 0, n |α|, this power series converges to 1/(1 − αz−1 ). Thus we have the z-transform pair

155

The z -Transform and Its Application to the Analysis of LTI Systems

Im(z)

x(n) |α| Re(z)

0 ROC

… 0 1 2 3 4 5



n

(a)

(b)

The exponential signal x(n) = α n u(n) (a), and the ROC of its ztransform (b).

Figure 1.2

1 , ROC: |z| > |α| (1.7) 1 − αz−1 The ROC is the exterior of a circle having radius |α|. Figure 1.2 shows a graph of the signal x(n) and its corresponding ROC. Note that, in general, α need not be real. If we set α = 1 in (1.7), we obtain the z-transform of the unit step signal z

x(n) = α n u(n) ←→ X(z) =

z

x(n) = u(n) ←→ X(z) =

1 , 1 − z−1

ROC: |z| > 1

(1.8)

EXAMPLE 1.4 Determine the z-transform of the signal



x(n) = −α n u(−n − 1) = Solution.

0, n≥0 −α n , n ≤ −1

From the definition (1.1) we have X(z) =

−1 

(−α n )z−n = −

n=−∞

∞  (α −1 z)l l=1

where l = −n. Using the formula A + A2 + A3 + · · · = A(1 + A + A2 + · · ·) = when |A| < 1 gives X(z) = −

A 1−A

1 α −1 z = 1 − α −1 z 1 − αz−1

provided that |α −1 z| < 1 or, equivalently, |z| < |α|. Thus z

x(n) = −α n u(−n − 1) ←→ X(z) = −

1 , 1 − αz−1

ROC: |z| < |α|

The ROC is now the interior of a circle having radius |α|. This is shown in Fig. 1.3.

156

(1.9)

The z -Transform and Its Application to the Analysis of LTI Systems

Im(z) x(n)



|α|

−5 −4 −3 −2 −1 0



n

Re(z)

ROC

0 |α|: In this case there is a ring in the z-plane where both power series converge simultaneously, as shown in Fig. 1.4(b). Then we obtain X(z) =

1 1 − 1 − αz−1 1 − bz−1

b−α = α + b − z − αbz−1

(1.10)

The ROC of X(z) is |α| < |z| < |b|.

This example shows that if there is a ROC for an infinite-duration two-sided signal, it is a ring (annular region) in the z -plane. From Examples 1.1, 1.3, 1.4, and 1.5, we see that the ROC of a signal depends both on its duration (finite or infinite) and on whether it is causal, anticausal, or two-sided. These facts are summarized in Table 1. One special case of a two-sided signal is a signal that has infinite duration on the right side but not on the left [i.e., x(n) = 0 for n < n0 < 0]. A second case is

158

The z -Transform and Its Application to the Analysis of LTI Systems

TABLE 1

Characteristic Families of Signals with Their Corresponding

ROCs Signal

ROC

Finite-Duration Signals Causal Entire z-plane except z = 0 0

n

0

n

0

n

Anticausal

Two-sided

Infinite-Duration Signals Causal r2 |z| > r2

… 0

n

Anticausal r1 |z| < r1

… 0

n

Two-sided r2 …

r2 < |z| < r1

… 0

n

r1

a signal that has infinite duration on the left side but not on the right [i.e., x(n) = 0 for n > n1 > 0]. A third special case is a signal that has finite duration on both the left and right sides [i.e., x(n) = 0 for n < n0 < 0 and n > n1 > 0]. These types of signals are sometimes called right-sided, left-sided, and finite-duration two-sided signals, respectively. The determination of the ROC for these three types of signals is left as an exercise for the reader (Problem 5). Finally, we note that the z-transform defined by (1.1) is sometimes referred to as the two-sided or bilateral z -transform, to distinguish it from the one-sided or

159

The z -Transform and Its Application to the Analysis of LTI Systems

unilateral z -transform given by X+ (z) =

∞ 

x(n)z−n

(1.11)

n=0

The one-sided z-transform is examined in Section 6. In this text we use the expression z-transform exclusively to mean the two-sided z-transform defined by (1.1). The term “two-sided” will be used only in cases where we want to resolve any ambiguities. Clearly, if x(n) is causal [i.e., x(n) = 0 for n < 0], the one-sided and two-sided z-transforms are identical. In any other case, they are different.

1.2

The Inverse z -Transform

Often, we have the z-transform X(z) of a signal and we must determine the signal sequence. The procedure for transforming from the z-domain to the time domain is called the inverse z -transform. An inversion formula for obtaining x(n) from X(z) can be derived by using the Cauchy integral theorem, which is an important theorem in the theory of complex variables. To begin, we have the z-transform defined by (1.1) as X(z) =

∞ 

x(k)z−k

(1.12)

k=−∞

Suppose that we multiply both sides of (1.12) by z n−1 and integrate both sides over a closed contour within the ROC of X(z) which encloses the origin. Such a contour is illustrated in Fig. 1.5. Thus we have 

 ˆ C

X(z)z

n−1

dz = ˆ C

∞ 

x(k)zn−1−k dz

(1.13)

k=−∞

where C denotes the closed contour in the ROC of X(z), taken in a counterclockwise direction. Since the series converges on this contour, we can interchange the order of Im(z)

r2

C Re(z) r1

Figure 1.5

Contour C for integral in (1.13).

160

The z -Transform and Its Application to the Analysis of LTI Systems

integration and summation on the right-hand side of (1.13). Thus (1.13) becomes   ∞  n−1 x(k) ˆ zn−1−k dz (1.14) ˆ X(z)z dz = C

k=−∞

C

Now we can invoke the Cauchy integral theorem, which states that   1 1, k = n n−1−k z dz = 0, k = n 2πj Cˆ

(1.15)

where C is any contour that encloses the origin. By applying (1.15), the right-hand side of (1.14) reduces to 2πj x(n) and hence the desired inversion formula  1 x(n) = X(z)zn−1 dz (1.16) 2πj Cˆ Although the contour integral in (1.16) provides the desired inversion formula for determining the sequence x(n) from the z-transform, we shall not use (1.16) directly in our evaluation of inverse z-transforms. In our treatment we deal with signals and systems in the z-domain which have rational z-transforms (i.e., z-transforms that are a ratio of two polynomials). For such z-transforms we develop a simpler method for inversion that stems from (1.16) and employs a table lookup.

2

Properties of the z -Transform The z-transform is a very powerful tool for the study of discrete-time signals and systems. The power of this transform is a consequence of some very important properties that the transform possesses. In this section we examine some of these properties. In the treatment that follows, it should be remembered that when we combine several z-transforms, the ROC of the overall transform is, at least, the intersection of the ROC of the individual transforms. This will become more apparent later, when we discuss specific examples. Linearity.

If z

x1 (n) ←→ X1 (z) and

z

x2 (n) ←→ X2 (z) then

z

x(n) = a1 x1 (n) + a2 x2 (n) ←→ X(z) = a1 X1 (z) + a2 X2 (z)

(2.1)

for any constants a1 and a2 . The proof of this property follows immediately from the definition of linearity and is left as an exercise for the reader. The linearity property can easily be generalized for an arbitrary number of signals. Basically, it implies that the z-transform of a linear combination of signals is the same linear combination of their z-transforms. Thus the linearity property helps us to find the z-transform of a signal by expressing the signal as a sum of elementary signals, for each of which, the z-transform is already known.

161

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 2.1 Determine the z-transform and the ROC of the signal x(n) = [3(2n ) − 4(3n )]u(n) Solution.

If we define the signals x1 (n) = 2n u(n)

and x2 (n) = 3n u(n) then x(n) can be written as x(n) = 3x1 (n) − 4x2 (n) According to (2.1), its z-transform is X(z) = 3X1 (z) − 4X2 (z) From (1.7) we recall that z

α n u(n) ←→

1 , 1 − αz−1

ROC: |z| > |α|

(2.2)

By setting α = 2 and α = 3 in (2.2), we obtain z

1 , 1 − 2z−1

ROC: |z| > 2

z

1 , 1 − 3z−1

ROC: |z| > 3

x1 (n) = 2n u(n) ←→ X1 (z) = x2 (n) = 3n u(n) ←→ X2 (z) =

The intersection of the ROC of X1 (z) and X2 (z) is |z| > 3. Thus the overall transform X(z) is X(z) =

3 4 − , −1 1 − 2z 1 − 3z−1

ROC: |z| > 3

EXAMPLE 2.2 Determine the z-transform of the signals (a) x(n) = (cos ω0 n)u(n) (b) x(n) = (sin ω0 n)u(n) Solution. (a) By using Euler’s identity, the signal x(n) can be expressed as x(n) = (cos ω0 n)u(n) =

1 j ω0 n 1 e u(n) + e−j ω0 n u(n) 2 2

Thus (2.1) implies that X(z) =

162

1 1 Z{ej ω0 n u(n)} + Z{e−j ω0 n u(n)} 2 2

The z -Transform and Its Application to the Analysis of LTI Systems

If we set α = e±j ω0 (|α| = |e±j ω0 | = 1) in (2.2), we obtain 1

z

ej ω0 n u(n) ←→

1 − ej ω0 z−1

,

ROC: |z| > 1

and z

e−j ω0 n u(n) ←→

1 , 1 − e−j ω0 z−1

ROC: |z| > 1

Thus X(z) =

1 1 1 1 + , j ω −1 −j 0 21−e z 2 1 − e ω0 z−1

ROC: |z| > 1

After some simple algebraic manipulations we obtain the desired result, namely, z

(cos ω0 n)u(n) ←→

1 − z−1 cos ω0 , 1 − 2z−1 cos ω0 + z−2

ROC: |z| > 1

(2.3)

(b) From Euler’s identity, x(n) = (sin ω0n)u(n) = Thus X(z) =

1 2j



1 1 − ej ω0 z−1



1 j ω0 n u(n) − e−j ω0 n u(n)] [e 2j 1



1 − e−j ω0 z−1

,

ROC: |z| > 1

and finally, z

(sin ω0 n)u(n) ←→

Time shifting.

z−1 sin ω0 , 1 − 2z−1 cos ω0 + z−2

ROC: |z| > 1

(2.4)

If z

x(n) ←→ X(z) then

z

x(n − k) ←→ z−k X(z)

(2.5)

The ROC of z−k X(z) is the same as that of X(z) except for z = 0 if k > 0 and z = ∞ if k < 0. The proof of this property follows immediately from the definition of the z-transform given in (1.1) The properties of linearity and time shifting are the key features that make the z-transform extremely useful for the analysis of discrete-time LTI systems. EXAMPLE 2.3 By applying the time-shifting property, determine the z-transform of the signals x2 (n) and x3 (n) in Example 1.1 from the z-transform of x1 (n).

163

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

It can easily be seen that x2 (n) = x1 (n + 2)

and x3 (n) = x1 (n − 2) Thus from (2.5) we obtain X2 (z) = z2 X1 (z) = z2 + 2z + 5 + 7z−1 + z−3 and

X3 (z) = z−2 X1 (z) = z−2 + 2z−3 + 5z−4 + 7z−5 + z−7

Note that because of the multiplication by z2 , the ROC of X2 (z) does not include the point z = ∞, even if it is contained in the ROC of X1 (z).

Example 2 .3 provides additional insight in understanding the meaning of the shifting property. Indeed, if we recall that the coefficient of z−n is the sample value at time n, it is immediately seen that delaying a signal by k(k > 0) samples [i.e., x(n) → x(n − k)] corresponds to multiplying all terms of the z-transform by z−k . The coefficient of z−n becomes the coefficient of z−(n+k) . EXAMPLE 2.4 Determine the transform of the signal  x(n) = Solution. Indeed,

1, 0,

0≤n≤N −1 elsewhere

(2.6)

We can determine the z-transform of this signal by using the definition

X(z) =

N−1 

 1·z

−n

=1+z

−1

+ ··· + z

−(N−1)

n=0

=

N, 1 − z−N , 1 − z−1

(1.1).

if z = 1 if z = 1

(2.7)

Since x(n) has finite duration, its ROC is the entire z-plane, except z = 0. Let us also derive this transform by using the linearity and time-shifting properties. Note that x(n) can be expressed in terms of two unit step signals x(n) = u(n) − u(n − N ) By using (2.1) and (2.5) we have X(z) = Z{u(n)} − Z{u(n − N )} = (1 − z−N )Z{u(n)} However, from (1.8) we have Z{u(n)} =

1 , 1 − z−1

which, when combined with (2.8), leads to (2.7).

164

ROC: |z| > 1

(2.8)

The z -Transform and Its Application to the Analysis of LTI Systems

Example 2.4 helps to clarify a very important issue regarding the ROC of the combination of several z-transforms. If the linear combination of several signals has finite duration, the ROC of its z-transform is exclusively dictated by the finiteduration nature of this signal, not by the ROC of the individual transforms. If

Scaling in the z -domain.

z

x(n) ←→ X(z), then

ROC: r1 < |z| < r2

z

a n x(n) ←→ X(a −1 z),

ROC: |a|r1 < |z| < |a|r2

(2.9)

for any constant a , real or complex.

Proof

From the definition (1.1) Z{a n x(n)} =

∞ 

a n x(n)z−n =

n=−∞

∞ 

x(n)(a −1 z)−n

n=−∞

= X(a −1 z) Since the ROC of X(z) is r1 < |z| < r2 , the ROC of X(a −1 z) is r1 < |a −1 z| < r2 or |a|r1 < |z| < |a|r2 To better understand the meaning and implications of the scaling property, we express a and z in polar form as a = r0 ej ω0 , z = rej ω , and we introduce a new complex variable w = a −1 z. Thus Z{x(n)} = X(z) and Z{a n x(n)} = X(w). It can easily be seen that   1 w = a −1 z = r ej (ω−ω0 ) r0 This change of variables results in either shrinking (if r0 > 1) or expanding (if r0 < 1) the z-plane in combination with a rotation (if ω0 = 2kπ ) of the z-plane (see Fig. 2.1). This explains why we have a change in the ROC of the new transform where |a| < 1. The case |a| = 1, that is, a = ej ω0 is of special interest because it corresponds only to rotation of the z-plane. z-plane Im(z)

Figure 2.1

Mapping of the z-plane to the w -plane via the transformation ω = a −1 z, a = r0 ej ω0 .

r 0

ω = a−1z

z

w-plane Im(w)

w ω − ω0

ω Re(z)

0

Re(w)

165

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 2.5 Determine the z-transforms of the signals (a) x(n) = a n (cos ω0 n)u(n) (b) x(n) = a n (sin ω0 n)u(n) Solution. (a) From (2.3) and (2.9) we easily obtain z

a n (cos ω0 n)u(n) ←→

1 − az−1 cos ω0 , 1 − 2az−1 cos ω0 + a 2 z−2

|z| > |a|

(2.10)

az−1 sin ω0 , 1 − 2az−1 cos ω0 + a 2 z−2

|z| > |a|

(2.11)

(b) Similarly, (2.4) and (2.9) yield z

a n (sin ω0 n)u(n) ←→

Time reversal.

If z

x(n) ←→ X(z),

ROC: r1 < |z| < r2

then z

x(−n) ←→ X(z−1 ),

Proof

ROC:

1 1 < |z| < r2 r1

(2.12)

From the definition (1.1), we have Z{x(−n)} =

∞ 

x(−n)z−n =

n=−∞

∞ 

x(l)(z−1 )−l = X(z−1 )

l=−∞

where the change of variable l = −n is made. The ROC of X(z−1 ) is r1 < |z−1 | < r2

or equivalently

1 1 < |z| < r2 r1

Note that the ROC for x(n) is the inverse of that for x(−n). This means that if z0 belongs to the ROC of x(n), then 1/z0 is in the ROC for x(−n). An intuitive proof of (2.12) is the following. When we fold a signal, the coefficient of z−n becomes the coefficient of zn . Thus, folding a signal is equivalent to replacing z by z−1 in the z-transform formula. In other words, reflection in the time domain corresponds to inversion in the z-domain. EXAMPLE 2.6 Determine the z-transform of the signal x(n) = u(−n)

166

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

It is known from (1.8) that z

u(n) ←→

1 , 1 − z−1

ROC: |z| > 1

1 , 1−z

ROC: |z| < 1

By using (2.12), we easily obtain z

u(−n) ←→

Differentiation in the z -domain.

(2.13)

If z

x(n) ←→ X(z) then z

nx(n) ←→ −z Proof

dX(z) dz

(2.14)

By differentiating both sides of (1.1), we have ∞ ∞   dX(z) −n−1 −1 = x(n)(−n)z = −z [nx(n)]z−n dz n=−∞ n=−∞

= −z−1 Z{nx(n)} Note that both transforms have the same ROC. EXAMPLE 2.7 Determine the z-transform of the signal x(n) = na n u(n) Solution. The signal x(n) can be expressed as nx1 (n), where x1 (n) = a n u(n). From (2.2) we have that 1 , 1 − az−1

ROC: |z| > |a|

az−1 dX1 (z) = , dz (1 − az−1 )2

ROC: |z| > |a|

z

x1 (n) = a n u(n) ←→ X1 (z) = Thus, by using (2.14), we obtain z

na n u(n) ←→ X(z) = −z

(2.15)

If we set a = 1 in (2.15), we find the z-transform of the unit ramp signal z

nu(n) ←→

z−1 , (1 − z−1 )2

ROC: |z| > 1

(2.16)

167

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 2.8 Determine the signal x(n) whose z-transform is given by X(z) = log(1 + az−1 ), Solution.

|z| > |a|

By taking the first derivative of X(z), we obtain dX(z) −az−2 = dz 1 + az−1

Thus −z

 dX(z) 1 = az−1 , dz 1 − (−a)z−1

|z| > |a|

The inverse z-transform of the term in brackets is (−a)n . The multiplication by z−1 implies a time delay by one sample (time-shifting property), which results in (−a)n−1 u(n − 1). Finally, from the differentiation property we have nx(n) = a(−a)n−1 u(n − 1) or x(n) = (−1)n+1

an u(n − 1) n

If

Convolution of two sequences.

z

x1 (n) ←→ X1 (z) z

x2 (n) ←→ X2 (z) then

z

x(n) = x1 (n) ∗ x2 (n) ←→ X(z) = X1 (z)X2 (z)

(2.17)

The ROC of X(z) is, at least, the intersection of that for X1 (z) and X2 (z). Proof The convolution of x1 (n) and x2 (n) is defined as x(n) =

∞ 

x1 (k)x2 (n − k)

k=−∞

The z-transform of x(n) is X(z) =

∞ 

x(n)z

−n

n=−∞

∞ 

=

n=−∞



∞ 

x1 (k)x2 (n − k) z−n

k=−∞

Upon interchanging the order of the summations and applying the time-shifting property in (2.5), we obtain

∞ ∞   x1 (k) x2 (n − k)z−n X(z) = n=−∞

k=−∞

= X2 (z)

∞  k=−∞

168

x1 (k)z−k = X2 (z)X1 (z)

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 2.9 Compute the convolution x(n) of the signals x1 (n) = {1, −2, 1}  1, 0 ≤ n ≤ 5 x2 (n) = 0, elsewhere Solution.

From (1.1), we have X1 (z) = 1 − 2z−1 + z−2 X2 (z) = 1 + z−1 + z−2 + z−3 + z−4 + z−5

According to (2.17), we carry out the multiplication of X1 (z) and X2 (z). Thus X(z) = X1 (z)X2 (z) = 1 − z−1 − z−6 + z−7 Hence x(n) = {1, −1, 0, 0, 0, 0, −1, 1} ↑

The same result can also be obtained by noting that X1 (z) = (1 − z−1 )2 X2 (z) = Then

1 − z−6 1 − z−1

X(z) = (1 − z−1 )(1 − z−6 ) = 1 − z−1 − z−6 + z−7

The reader is encouraged to obtain the same result explicitly by using the convolution summation formula (time-domain approach).

The convolution property is one of the most powerful properties of the z-transform because it converts the convolution of two signals (time domain) to multiplication of their transforms. Computation of the convolution of two signals, using the z-transform, requires the following steps: 1. Compute the z-transforms of the signals to be convolved. X1 (z) = Z{x1 (n)} (time domain −→ z-domain) X2 (z) = Z{x2 (n)} 2. Multiply the two z-transforms. X(z) = X1 (z)X2 (z),

(z-domain)

3. Find the inverse z-transform of X(z). x(n) = Z −1 {X(z)},

(z-domain −→ time domain)

169

The z -Transform and Its Application to the Analysis of LTI Systems

This procedure is, in many cases, computationally easier than the direct evaluation of the convolution summation. Correlation of two sequences.

If z

x1 (n) ←→ X1 (z) z

x2 (n) ←→ X2 (z) then rx1 x2 (l) =

∞ 

z

x1 (n)x2 (n − l) ←→ Rx1 x2 (z) = X1 (z)X2 (z−1 )

(2.18)

n=−∞

Proof

We recall that rx1 x2 (l) = x1 (l) ∗ x2 (−l)

Using the convolution and time-reversal properties, we easily obtain Rx1 x2 (z) = Z{x1 (l)}Z{x2 (−l)} = X1 (z)X2 (z−1 ) The ROC of Rx1 x2 (z) is at least the intersection of that for X1 (z) and X2 (z−1 ). As in the case of convolution, the crosscorrelation of two signals is more easily done via polynomial multiplication according to (2.18) and then inverse transforming the result. EXAMPLE 2.10 Determine the autocorrelation sequence of the signal x(n) = a n u(n), Solution. gives

−1 < a < 1

Since the autocorrelation sequence of a signal is its correlation with itself, (2.18) Rxx (z) = Z{rxx (l)} = X(z)X(z−1 )

From (2.2) we have X(z) =

1 , 1 − az−1

ROC: |z| > |a|

(causal signal)

and by using (2.15), we obtain X(z−1 ) =

1 , 1 − az

ROC: |z|
0 and z = ∞ if k < 0

Scaling in the z-domain

a n x(n)

X(a −1 z)

|a|r2 < |z| < |a|r1

Time reversal

x(−n)

X(z−1 )

1 r1

Conjugation

x ∗ (n)

X ∗ (z∗ )

ROC

Real part

Re{x(n)}

1 [X(z) 2

Imaginary part

Im{x(n)}

1 j [X(z) 2

Differentiation in the z-domain

nx(n)

−z dX(z) dz

r2 < |z| < r1

Convolution

x1 (n) ∗ x2 (n)

X1 (z)X2 (z)

At least, the intersection of ROC 1 and ROC 2

Correlation

rx1 x2 (l) = x1 (l) ∗ x2 (−l)

Rx1 x2 (z) = X1 (z)X2 (z−1 )

At least, the intersection of ROC of X1 (z) and X2 (z−1 )

Initial value theorem Multiplication

If x(n) causal

Parseval’s relation

+ X ∗ (z∗ )] − X ∗ (z∗ )]

x(0) = lim X(z) z→∞   z  −1 1 x1 (n)x2 (n) 2πj C ˆ X1 (v)X2 v v dv  ∞  1 ∗ ∗ −1 x1 (n)x2∗ (n) = 2πj ˆ X1 (v)X2 (1/v )v dv

n=−∞

< |z|
1

3

a n u(n)

1 1 − z−1 1 1 − az−1

4

na n u(n)

az−1 (1 − az−1 )2

|z| > |a|

5

−a n u(−n − 1)

1 1 − az−1

|z| < |a|

6

−na n u(−n − 1)

az−1 (1 − az−1 )2

|z| < |a|

7

(cos ω0 n)u(n)

1 − z−1 cos ω0 1 − 2z−1 cos ω0 + z−2

|z| > 1

8

(sin ω0 n)u(n)

z−1 sin ω0 1 − 2z−1 cos ω0 + z−2

|z| > 1

9

(a n cos ω0 n)u(n)

1 − az−1 cos ω0 1 − 2az−1 cos ω0 + a 2 z−2

|z| > |a|

10

(a n sin ω0 n)u(n)

az−1 sin ω0 1 − 2az−1 cos ω0 + a 2 z−2

|z| > |a|

|z| > |a|

Rational z -Transforms As indicated in Section 2, an important family of z-transforms are those for which X(z) is a rational function, that is, a ratio of two polynomials in z−1 (or z). In this section we discuss some very important issues regarding the class of rational z-transforms.

3.1

Poles and Zeros

The zeros of a z-transform X(z) are the values of z for which X(z) = 0. The poles of a z-transform are the values of z for which X(z) = ∞. If X(z) is a rational function, then M −k B(z) b0 + b1 z−1 + · · · + bM z−M k=0 bk z X(z) = = = (3.1) N −1 −N −k A(z) a0 + a 1 z + · · · + a N z k=0 ak z  0, we can avoid the negative powers of z by factoring out the If a0 = 0 and b0 = terms b0 z−M and a0 z−N as follows: X(z) =

174

b0 z−M zM + (b1 /b0 )zM−1 + · · · + bM /b0 B(z) = A(z) a0 z−N zN + (a1 /a0 )zN−1 + · · · + aN /a0

The z -Transform and Its Application to the Analysis of LTI Systems

Since B(z) and A(z) are polynomials in z, they can be expressed in factored form as X(z) =

(z − z1 )(z − z2 ) · · · (z − zM ) b0 B(z) = z−M+N A(z) a0 (z − p1 )(z − p2 ) · · · (z − pN ) M 

X(z) =

(z − zk )

(3.2)

k=1 GzN−M N 

(z − pk )

k=1

where G ≡ b0 /a0 . Thus X(z) has M finite zeros at z = z1 , z2 , . . . , zM (the roots of the numerator polynomial), N finite poles at z = p1 , p2 , . . . , pN (the roots of the denominator polynomial), and |N − M| zeros (if N > M ) or poles (if N < M ) at the origin z = 0. Poles or zeros may also occur at z = ∞. A zero exists at z = ∞ if X(∞) = 0 and a pole exists at z = ∞ if X(∞) = ∞. If we count the poles and zeros at zero and infinity, we find that X(z) has exactly the same number of poles as zeros. We can represent X(z) graphically by a pole–zero plot (or pattern) in the complex plane, which shows the location of poles by crosses (×) and the location of zeros by circles (◦). The multiplicity of multiple-order poles or zeros is indicated by a number close to the corresponding cross or circle. Obviously, by definition, the ROC of a z-transform should not contain any poles. EXAMPLE 3.1 Determine the pole–zero plot for the signal x(n) = a n u(n), Solution.

a>0

From Table 3 we find that X(z) =

1 z = , 1 − az−1 z−a

ROC: |z| > a

Thus X(z) has one zero at z1 = 0 and one pole at p1 = a . The pole–zero plot is shown in Fig. 1. Note that the pole p 1 = a is not included in the ROC since the z-transform does not converge at a pole. Im(z)

ROC a 0

Re(z)

Figure 3.1

Pole–zero plot for the causal exponential signal x(n) = a n u(n).

175

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 3.2 Determine the pole–zero plot for the signal  x(n) =

an, 0,

0≤n≤M −1 elsewhere

where a > 0. Solution.

From the definition (1.1) we obtain

X(z) =

M−1 

(az−1 )n =

n=0

1 − (az−1 )M zM − a M = M−1 −1 1 − az z (z − a)

Since a > 0, the equation zM = a M has M roots at zk = aej 2π k/M

k = 0, 1, . . . , M − 1

The zero z0 = a cancels the pole at z = a . Thus

X(z) =

(z − z1 )(z − z2 ) · · · (z − zM−1 ) zM−1

which has M − 1 zeros and M − 1 poles, located as shown in Fig 3.2 for M = 8. Note that the ROC is the entire z-plane except z = 0 because of the M − 1 poles located at the origin.

Im(z)

M −1 poles

Figure 3.2

Pole–zero pattern for the finite-duration signal x(n) = a n , 0 ≤ n ≤ M − 1(a > 0), for M = 8.

176

|z| = a

Re(z)

The z -Transform and Its Application to the Analysis of LTI Systems

Im(z)

ROC p1 z1 r

ω0 z2 ω0

Re(z) p2

Figure 3.3

Pole-zero pattern for Example 3.3.

Clearly, if we are given a pole–zero plot, we can determine X(z), by using (3.2), to within a scaling factor G. This is illustrated in the following example. EXAMPLE 3.3 Determine the z-transform and the signal that corresponds to the pole–zero plot of Fig. 3.3. Solution. There are two zeros (M = 2) at z1 = 0, z2 = r cos ω0 and two poles (N = 2) at p1 = rej ω0 , p2 = re−j ω0 . By substitution of these relations into (3.2), we obtain X(z) = G

z(z − r cos ω0) (z − z1 )(z − z2 ) =G ), (z − p1 )(z − p2 ) (z − rej ω0 )(z − re−j ω0

ROC: |z| > r

After some simple algebraic manipulations, we obtain X(z) = G

1 − rz−1 cos ω0 , 1 − 2rz−1 cos ω0 + r 2 z−2

ROC: |z| > r

From Table 3 we find that x(n) = G(r n cos ω0 n)u(n)

From Example 3.3, we see that the product (z − p1 )(z − p2 ) results in a polynomial with real coefficients, when p1 and p2 are complex conjugates. In general, if a polynomial has real coefficients, its roots are either real or occur in complexconjugate pairs. As we have seen, the z-transform X(z) is a complex function of the complex variable z = (z) + j (z). Obviously, |X(z)|, the magnitude of X(z), is a real and positive function of z. Since z represents a point in the complex plane, |X(z)| is a two-dimensional function and describes a “surface.” This is illustrated in Fig. 3.4 for the z-transform z−1 − z−2 X(z) = (3.3) 1 − 1.2732z−1 + 0.81z−2 which has one zero at z1 = 1 and two poles at p1 , p2 = 0.9e±j π/4 . Note the high peaks near the singularities (poles) and the deep valley close to the zero.

177

The z -Transform and Its Application to the Analysis of LTI Systems

Figure 3.4

3.2

Graph of |X(z)| for the z-transform in (3.3).

Pole Location and Time-Domain Behavior for Causal Signals

In this subsection we consider the relation between the z-plane location of a pole pair and the form (shape) of the corresponding signal in the time domain. The discussion is based generally on the collection of z-transform pairs given in Table 3 and the results in the preceding subsection. We deal exclusively with real, causal signals. In particular, we see that the characteristic behavior of causal signals depends on whether the poles of the transform are contained in the region |z| < 1, or in the region |z| > 1, or on the circle |z| = 1. Since the circle |z| = 1 has a radius of 1, it is called the unit circle. If a real signal has a z-transform with one pole, this pole has to be real. The only such signal is the real exponential z

x(n) = a n u(n) ←→ X(z) =

1 , 1 − az−1

ROC: |z| > |a|

having one zero at z1 = 0 and one pole at p1 = a on the real axis. Figure 3.5 illustrates the behavior of the signal with respect to the location of the pole relative to the unit circle. The signal is decaying if the pole is inside the unit circle, fixed if the pole is on the unit circle, and growing if the pole is outside the unit circle. In addition, a negative pole results in a signal that alternates in sign. Obviously, causal signals with poles outside the unit circle become unbounded, cause overflow in digital systems, and in general, should be avoided. A causal real signal with a double real pole has the form x(n) = na n u(n) (see Table 3) and its behavior is illustrated in Fig. 3.6. Note that in contrast to the single-pole signal, a double real pole on the unit circle results in an unbounded signal.

178

The z -Transform and Its Application to the Analysis of LTI Systems

x(n)

z-plane

0

1

… n

0

0

x(n)

z-plane

x(n)

z-plane

1

… n

0

x(n)

z-plane …

0

1

0

n

… 0

x(n)

z-plane

1

0

n

x(n)

z-plane …

0

1

0

n

… 0

1

0

n

Figure 3.5 Time-domain behavior of a single-real-pole causal signal as a function of

the location of the pole with respect to the unit circle. x(n)

z-plane m=2 0

1

z-plane … n

0

m=2 0

x(n)

z-plane

x(n)

1

z-plane

… n

0

x(n)

… 0

m=2 1

0

m=2 1

m=2 0

x(n)

z-plane

0

n

… 1

0

x(n)

z-plane



n



m=2 0

n

0

1

0

n

Figure 3.6 Time-domain behavior of causal signals corresponding to a double (m = 2) real pole, as a function of the pole location.

179

The z -Transform and Its Application to the Analysis of LTI Systems

x(n)

z-plane

rn r ωo 0

1

0

n

x(n)

z-plane

r=1

1 ωo 0

1

0

n

x(n)

z-plane

rn

r ωo 0

1

0

n

Figure 3.7 A pair of complex-conjugate poles corresponds to causal

signals with oscillatory behavior.

Figure 3.7 illustrates the case of a pair of complex-conjugate poles. According to Table 3, this configuration of poles results in an exponentially weighted sinusoidal signal. The distance r of the poles from the origin determines the envelope of the sinusoidal signal and their angle with the real positive axis, its relative frequency. Note that the amplitude of the signal is growing if r > 1, constant if r = 1 (sinusoidal signals), and decaying if r < 1. Finally, Fig. 3.8 shows the behavior of a causal signal with a double pair of poles on the unit circle. This reinforces the corresponding results in Fig. 3.6 and illustrates that multiple poles on the unit circle should be treated with great care. To summarize, causal real signals with simple real poles or simple complexconjugate pairs of poles, which are inside or on the unit circle, are always bounded in amplitude. Furthermore, a signal with a pole (or a complex-conjugate pair of poles)

180

The z -Transform and Its Application to the Analysis of LTI Systems

x(n)

z-plane m=2 ω0 0

1

0

n

m=2

Figure 3.8 Causal signal corresponding to a double pair of

complex-conjugate poles on the unit circle.

near the origin decays more rapidly than one associated with a pole near (but inside) the unit circle. Thus the time behavior of a signal depends strongly on the location of its poles relative to the unit circle. Zeros also affect the behavior of a signal but not as strongly as poles. For example, in the case of sinusoidal signals, the presence and location of zeros affects only their phase. At this point, it should be stressed that everything we have said about causal signals applies as well to causal LTI systems, since their impulse response is a causal signal. Hence if a pole of a system is outside the unit circle, the impulse response of the system becomes unbounded and, consequently, the system is unstable.

3.3

The System Function of a Linear Time-Invariant System

Recall that the output of a (relaxed) linear time-invariant system to an input sequence x(n) can be obtained by computing the convolution of x(n) with the unit sample response of the system. The convolution property, derived in Section 2, allows us to express this relationship in the z-domain as Y (z) = H (z)X(z)

(3.4)

where Y (z) is the z-transform of the output sequence y(n), X(z) is the z-transform of the input sequence x(n) and H (z) is the z-transform of the unit sample response h(n). If we know h(n) and x(n), we can determine their corresponding z-transforms H (z) and X(z), multiply them to obtain Y (z), and therefore determine y(n) by evaluating the inverse z-transform of Y (z). Alternatively, if we know x(n) and we observe the output y(n) of the system, we can determine the unit sample response by first solving for H (z) from the relation H (z) =

Y (z) X(z)

(3.5)

and then evaluating the inverse z-transform of H (z).

181

The z -Transform and Its Application to the Analysis of LTI Systems

Since

∞ 

H (z) =

h(n)z−n

(3.6)

n=−∞

it is clear that H (z) represents the z-domain characterization of a system, whereas h(n) is the corresponding time-domain characterization of the system. In other words, H (z) and h(n) are equivalent descriptions of a system in the two domains. The transform H (z) is called the system function. The relation in (3.5) is particularly useful in obtaining H (z) when the system is described by a linear constant-coefficient difference equation of the form y(n) = −

N 

ak y(n − k) +

M 

bk x(n − k)

(3.7)

k=0

k=1

In this case the system function can be determined directly from (3.7) by computing the z-transform of both sides of (3.7). Thus, by applying the time-shifting property, we obtain Y (z) = −

N 

ak Y (z)z−k +

Y (z) 1 +

N 

 ak z−k

= X(z)

M 

 bk z−k

k=0

k=1

bk X(z)z−k

k=0

k=1



M 

(3.8) M 

Y (z) = H (z) = X(z)

bk z−k

k=0 N 

1+

ak z−k

k=1

Therefore, a linear time-invariant system described by a constant-coefficient difference equation has a rational system function. This is the general form for the system function of a system described by a linear constant-coefficient difference equation. From this general form we obtain two important special forms. First, if ak = 0 for 1 ≤ k ≤ N , (3.8) reduces to H (z) =

M  k=0

bk z−k =

M 1  bk zM−k zM

(3.9)

k=0

In this case, H (z) contains M zeros, whose values are determined by the system parameters {bk }, and an M th-order pole at the origin z = 0. Since the system contains only trivial poles (at z = 0) and M nontrivial zeros, it is called an all-zero system. Clearly, such a system has a finite-duration impulse response (FIR), and it is called an FIR system or a moving average (MA) system.

182

The z -Transform and Its Application to the Analysis of LTI Systems

On the other hand, if bk = 0 for 1 ≤ k ≤ M , the system function reduces to H (z) =

1+

b0 N

k=1 ak z

−k

= N

b0 zN

k=0

ak zN−k

,

a0 ≡ 1

(3.10)

In this case H (z) consists of N poles, whose values are determined by the system parameters {ak } and an N th-order zero at the origin z = 0. We usually do not make reference to these trivial zeros. Consequently, the system function in (3.10) contains only nontrivial poles and the corresponding system is called an all-pole system. Due to the presence of poles, the impulse response of such a system is infinite in duration, and hence it is an IIR system. The general form of the system function given by (3.8) contains both poles and zeros, and hence the corresponding system is called a pole–zero system, with N poles and M zeros. Poles and/or zeros at z = 0 and z = ∞ are implied but are not counted explicitly. Due to the presence of poles, a pole–zero system is an IIR system. The following example illustrates the procedure for determining the system function and the unit sample response from the difference equation. EXAMPLE 3.4 Determine the system function and the unit sample response of the system described by the difference equation 1 y(n) = y(n − 1) + 2x(n) 2 Solution.

By computing the z-transform of the difference equation, we obtain Y (z) =

1 −1 z Y (z) + 2X(z) 2

H (z) =

Y (z) 2 = X(z) 1 − 21 z−1

Hence the system function is

This system has a pole at z = transform

1 2

and a zero at the origin. Using Table 3 we obtain the inverse 1 h(n) = 2( )n u(n) 2

This is the unit sample response of the system.

We have now demonstrated that rational z-transforms are encountered in commonly used systems and in the characterization of linear time-invariant systems. In Section 4 we describe several methods for determining the inverse z-transform of rational functions.

183

The z -Transform and Its Application to the Analysis of LTI Systems

4 Inversion of the z -Transform As we saw in Section 1.2, the inverse z-transform is formally given by x(n) =

 1 X(z)zn−1 dz 2πj Cˆ

(4.1)

where the integral is a contour integral over a closed path C that encloses the origin and lies within the region of convergence of X(z). For simplicity, C can be taken as a circle in the ROC of X(z) in the z-plane. There are three methods that are often used for the evaluation of the inverse z-transform in practice: 1. Direct evaluation of (4.1), by contour integration. 2. Expansion into a series of terms, in the variables z, and z−1 . 3. Partial-fraction expansion and table lookup.

4.1

The Inverse z -Transform by Contour Integration

In this section we demonstrate the use of the Cauchy’s integral theorem to determine the inverse z-transform directly from the contour integral. Let f (z) be a function of the complex variable z and C be a closed path in the z-plane. If the derivative df (z)/dz exists on and inside the contour C and if f (z) has no poles at z = z0 , then

Cauchy’s integral theorem.

  f (z) 1 f (z0 ), dz = 0, 2πj Cˆ z − z0

if z0 is inside C if z0 is outside C

(4.2)

More generally, if the (k + 1)-order derivative of f (z) exists and f (z) has no poles at z = z0 , then  

 k−1  d f (z) 1  f (z) 1 , if z0 is inside C k−1  (k − 1)! dz = dz z=z0  2πj Cˆ (z − z0 )k 0, if z0 is outside C 

(4.3)

The values on the right-hand side of (4.2) and (4.3) are called the residues of the pole at z = z0 . The results in (4.2) and (4.3) are two forms of the Cauchy’s integral theorem. We can apply (4.2) and (4.3) to obtain the values of more general contour integrals. To be specific, suppose that the integrand of the contour integral is a

184

The z -Transform and Its Application to the Analysis of LTI Systems

proper fraction f (z)/g(z), where f (z) has no poles inside the contour C and g(z) is a polynomial with distinct (simple) roots z1 , z2 , . . . , zn inside C . Then    n f (z) Ai 1 1 dz dz = 2πj Cˆ g(z) 2πj Cˆ z − zi i=1

=

n  i=1

=

n 



1 Ai dz 2πj Cˆ z − zi

(4.4)

Ai

i=1

 f (z)  Ai = (z − zi ) g(z) z=zi

where

(4.5)

The values {Ai } are residues of the corresponding poles at z = zi , i = 1, 2, . . . , n. Hence the value of the contour integral is equal to the sum of the residues of all the poles inside the contour C . We observe that (4.4) was obtained by performing a partial-fraction expansion of the integrand and applying (4.2). When g(z) has multiple-order roots as well as simple roots inside the contour, the partial-fraction expansion, with appropriate modifications, and (4.3) can be used to evaluate the residues at the corresponding poles. In the case of the inverse z-transform, we have  1 X(z)zn−1 dz x(n) = 2πj Cˆ  = [residue of X(z)zn−1 at z = zi ] (4.6) all poles {zi } inside C

=



(z − zi )X(z)zn−1 |z=zi

i

provided that the poles {zi } are simple. If X(z)zn−1 has no poles inside the contour C for one or more values of n, then x(n) = 0 for these values. The following example illustrates the evaluation of the inverse z-transform by use of the Cauchy’s integral theorem. EXAMPLE 4.1 Evaluate the inverse z-transform of X(z) =

1 , 1 − az−1

|z| > |a|

using the complex inversion integral.

185

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

We have x(n) =

  n zn−1 1 z dz 1 dz = 2πj Cˆ 1 − az−1 2πj Cˆ z − a

where C is a circle at radius greater than |a|. We shall evaluate this integral using (4.2) with f (z) = zn . We distinguish two cases. 1. If n ≥ 0, f (z) has only zeros and hence no poles inside C . The only pole inside C is z = a . Hence n≥0 x(n) = f (z0 ) = a n , 2. If n < 0, f (z) = zn has an nth-order pole at z = 0, which is also inside C . Thus there are contributions from both poles. For n = −1 we have    1 1  1  1 + =0 dz = x(−1) = 2πj Cˆ z(z − a) z − a z=0 z z=a If n = −2, we have x(−2) =

      d 1 1 1  + 1 dz = =0  ˆ 2 2 2πj C z (z − a) dz z − a z=0 z z=a

By continuing in the same way we can show that x(n) = 0 for n < 0. Thus x(n) = a n u(n)

4.2

The Inverse z -Transform by Power Series Expansion

The basic idea in this method is the following: Given a z-transform X(z) with its corresponding ROC, we can expand X(z) into a power series of the form X(z) =

∞ 

cn z−n

(4.7)

n=−∞

which converges in the given ROC. Then, by the uniqueness of the z-transform, x(n) = cn for all n. When X(z) is rational, the expansion can be performed by long division. To illustrate this technique, we will invert some z-transforms involving the same expression for X(z), but different ROC. This will also serve to emphasize again the importance of the ROC in dealing with z-transforms. EXAMPLE 4.2 Determine the inverse z-transform of X(z) = when (a) ROC: |z| > 1 (b) ROC: |z| < 0.5

186

1 1 − 1.5z−1 + 0.5z−2

The z -Transform and Its Application to the Analysis of LTI Systems

Solution. (a) Since the ROC is the exterior of a circle, we expect x(n) to be a causal signal. Thus we seek a power series expansion in negative powers of z. By dividing the numerator of X(z) by its denominator, we obtain the power series X(z) =

1 1 − 23 z−1 + 21 z−2

3 7 15 31 = 1 + z−1 + z−2 + z−3 + z−4 + · · · 2 4 8 16

By comparing this relation with (1.1), we conclude that 3 7 15 31 x(n) = {1, , , , , . . .} ↑ 2 4 8 16 Note that in each step of the long-division process, we eliminate the lowest-power term of z−1 . (b) In this case the ROC is the interior of a circle. Consequently, the signal x(n) is anticausal. To obtain a power series expansion in positive powers of z, we perform the long division in the following way: 2 3 4 5 6

2z + 6z + 14z + 30z + 62z + · · · 1 −2 3 −1 z − 2z + 1 1 2 1 − 3z + 2z2 3z − 2z2 3z − 9z2 + 6z3 7z2 − 6z3 7z2 − 21z3 + 14z4 15z3 − 14z4 15z3 − 45z4 + 30z5 31z4 − 30z5 Thus X(z) =

1 1 − 23 z−1 + 21 z−2

= 2z2 + 6z3 + 14z4 + 30z5 + 62z6 + · · ·

In this case x(n) = 0 for n ≥ 0. By comparing this result to (1.1), we conclude that x(n) = {· · · 62, 30, 14, 6, 2, 0, 0} ↑

We observe that in each step of the long-division process, the lowest-power term of z is eliminated. We emphasize that in the case of anticausal signals we simply carry out the long division by writing down the two polynomials in “reverse” order (i.e., starting with the most negative term on the left).

From this example we note that, in general, the method of long division will not provide answers for x(n) when n is large because the long division becomes tedious. Although the method provides a direct evaluation of x(n), a closed-form solution is not possible, except if the resulting pattern is simple enough to infer the general term x(n). Hence this method is used only if one wishes to determine the values of the first few samples of the signal.

187

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 4.3 Determine the inverse z-transform of X(z) = log(1 + az−1 ), Solution.

|z| > |a|

Using the power series expansion for log(1 + x), with |x| < 1, we have X(z) =

∞  (−1)n+1 a n z−n n n=1

Thus

 x(n) =

n

(−1)n+1 an , n ≥ 1 0, n≤0

Expansion of irrational functions into power series can be obtained from tables.

4.3

The Inverse z -Transform by Partial-Fraction Expansion

In the table lookup method, we attempt to express the function X(z) as a linear combination X(z) = α1 X1 (z) + α2 X2 (z) + · · · + αK XK (z) (4.8) where X1 (z), . . . , XK (z) are expressions with inverse transforms x1 (n), . . . , xK (n) available in a table of z-transform pairs. If such a decomposition is possible, then x(n), the inverse z-transform of X(z), can easily be found using the linearity property as (4.9) x(n) = α1 x1 (n) + α2 x2 (n) + · · · + αK xK (n) This approach is particularly useful if X (z) is a rational function, as in (3.1). Without loss of generality, we assume that a0 = 1, so that (3.1) can be expressed as X(z) =

b0 + b1 z−1 + · · · + bM z−M B(z) = A(z) 1 + a1 z−1 + · · · + aN z−N

(4.10)

Note that if a0 = 1, we can obtain (4.10) from (3.1) by dividing both numerator and denominator by a0 . A rational function of the form (4.10) is called proper if aN = 0 and M < N . From (3.2) it follows that this is equivalent to saying that the number of finite zeros is less than the number of finite poles. An improper rational function (M ≥ N ) can always be written as the sum of a polynomial and a proper rational function. This procedure is illustrated by the following example. EXAMPLE 4.4 Express the improper rational transform X(z) =

1 + 3z−1 + 1+

11 −2 z + 13 z−3 6 5 −1 z + 16 z−2 6

in terms of a polynomial and a proper function.

188

The z -Transform and Its Application to the Analysis of LTI Systems

Solution. First, we note that we should reduce the numerator so that the terms z−2 and z−3 are eliminated. Thus we should carry out the long division with these two polynomials written in reverse order. We stop the division when the order of the remainder becomes z−1 . Then we obtain 1 −1 z 6 X(z) = 1 + 2z−1 + 5 −1 1 + 6 z + 16 z−2

In general, any improper rational function (M ≥ N ) can be expressed as X(z) =

B(z) B1 (z) = c0 + c1 z−1 + · · · + cM−N z−(M−N) + A(z) A(z)

(4.11)

The inverse z-transform of the polynomial can easily be found by inspection. We focus our attention on the inversion of proper rational transforms, since any improper function can be transformed into a proper function by using (4.11). We carry out the development in two steps. First, we perform a partial fraction expansion of the proper rational function and then we invert each of the terms. Let X(z) be a proper rational function, that is, X(z) =

b0 + b1 z−1 + · · · + bM z−M B(z) = A(z) 1 + a1 z−1 + · · · + aN z−N

(4.12)

where aN = 0

and

M M , the function X(z) b0 zN−1 + b1 zN−2 + · · · + bM zN−M−1 = z zN + a1 zN−1 + · · · + aN

(4.14)

is also always proper. Our task in performing a partial-fraction expansion is to express (4.14) or, equivalently, (4.12) as a sum of simple fractions. For this purpose we first factor the denominator polynomial in (4.14) into factors that contain the poles p1 , p2 , . . . , pN of X(z). We distinguish two cases. Suppose that the poles p1 , p2 , . . . , pN are all different (distinct). Then we seek an expansion of the form

Distinct poles.

X(z) A1 A2 AN = + + ··· + z z − p1 z − p2 z − pN

(4.15)

The problem is to determine the coefficients A1 , A2 , . . . , AN . There are two ways to solve this problem, as illustrated in the following example.

189

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 4.5 Determine the partial-fraction expansion of the proper function X(z) =

1 1 − 1.5z−1 + 0.5z−2

(4.16)

Solution. First we eliminate the negative powers, by multiplying both numerator and denominator by z2 . Thus z2 X(z) = 2 z − 1.5z + 0.5 The poles of X(z) are p1 = 1 and p2 = 0.5. Consequently, the expansion of the form (4.15) is X(z) z A1 A2 = = + z (z − 1)(z − 0.5) z − 1 z − 0.5

(4.17)

A very simple method to determine A1 and A2 is to multiply the equation by the denominator term (z − 1)(z − 0.5). Thus we obtain z = (z − 0.5)A1 + (z − 1)A2

(4.18)

Now if we set z = p1 = 1 in (4.18), we eliminate the term involving A2 . Hence 1 = (1 − 0.5)A1 Thus we obtain the result A1 = 2. Next we return to (4.18) and set z = p 2 = 0.5, thus eliminating the term involving A1 , so we have 0.5 = (0.5 − 1)A2 and hence A2 = −1. Therefore, the result of the partial-fraction expansion is X(z) 2 1 = − z z − 1 z − 0.5

(4.19)

The example given above suggests that we can determine the coefficients A1 , A2 , . . . , AN , by multiplying both sides of (4.15) by each of the terms (z − pk ), k = 1, 2, . . . , N , and evaluating the resulting expressions at the corresponding pole positions, p1 , p2 , . . . , pN . Thus we have, in general, (z − pk )X(z) (z − pk )A1 (z − pk )AN = + · · · + Ak + · · · + z z − p1 z − pN

(4.20)

Consequently, with z = pk , (4.20) yields the kth coefficient as  (z − pk )X(z)  Ak = ,  z z=pk

190

k = 1, 2, . . . , N

(4.21)

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 4.6 Determine the partial-fraction expansion of X(z) =

1 + z−1 1 − z−1 + 0.5z−2

(4.22)

Solution. To eliminate negative powers of z in (4.22), we multiply both numerator and denominator by z 2 . Thus z+1 X(z) = 2 z z − z + 0.5 The poles of X(z) are complex conjugates p1 =

1 1 +j 2 2

p2 =

1 1 −j 2 2

and

Since p1 = p2 , we seek an expansion of the form (4.15). Thus X(z) z+1 A1 A2 = = + z (z − p1 )(z − p2 ) z − p1 z − p2 To obtain A1 and A2 , we use the formula (4.21). Thus we obtain

A1 =

  (z − p1 )X(z)  z + 1  = =  z z − p2 z=p1 z=p1

 (z − p2 )X(z)  A2 =  z

z=p2

 z + 1  = = z − p1 z=p2

1 2 1 2

+j − 1 2

1 2 1 2

+ j 21 + 1 1 2

+j

1 2

− j 21 + 1

− j 21 −

1 2

− j 21

=

1 3 −j 2 2

=

1 3 +j 2 2

The expansion (4.15) and the formula (4.21) hold for both real and complex poles. The only constraint is that all poles be distinct. We also note that A2 = A∗1 . It can be easily seen that this is a consequence of the fact that p2 = p1∗ . In other words, complex-conjugate poles result in complex-conjugate coefficients in the partialfraction expansion. This simple result will prove very useful later in our discussion.

If X(z) has a pole of multiplicity l , that is, it contains in its denominator the factor (z − pk )l , then the expansion (4.15) is no longer true. In this case a different expansion is needed. First, we investigate the case of a double pole (i.e., l = 2).

Multiple-order poles.

191

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 4.7 Determine the partial-fraction expansion of X(z) = Solution.

1 (1 + z−1 )(1 − z−1 )2

(4.23)

First, we express (4.23) in terms of positive powers of z, in the form X(z) z2 = z (z + 1)(z − 1)2

X(z) has a simple pole at p1 = −1 and a double pole p2 = p3 = 1. In such a case the appropriate partial-fraction expansion is X(z) z2 A1 A2 A3 = = + + z (z + 1)(z − 1)2 z + 1 z − 1 (z − 1)2

(4.24)

The problem is to determine the coefficients A1 , A2 , and A3 . We proceed as in the case of distinct poles. To determine A1 , we multiply both sides of (4.24) by (z + 1) and evaluate the result at z = −1. Thus (4.24) becomes (z + 1)X(z) z+1 z+1 = A1 + A3 A2 + z z−1 (z − 1)2 which, when evaluated at z = −1, yields A1 =

 (z + 1)X(z)  1 =  z 4 z=−1

Next, if we multiply both sides of (4.24) by (z − 1)2 , we obtain (z − 1)2 X(z) (z − 1)2 = A1 + (z − 1)A2 + A3 z z+1

(4.25)

Now, if we evaluate (4.25) at z = 1, we obtain A3 . Thus  (z − 1)2X(z)  1 = A3 =  z 2 z=1 The remaining coefficient A2 can be obtained by differentiating both sides of (4.25) with respect to z and evaluating the result at z = 1. Note that it is not necessary formally to carry out the differentiation of the right-hand side of (4.25), since all terms except A2 vanish when we set z = 1. Thus  d (z − 1)2 X(z) 3 = (4.26) A2 = dz z 4 z=1

The generalization of the procedure in the example above to the case of an mthorder pole (z − pk )m is straightforward. The partial-fraction expansion must contain the terms A1k A2k Amk + + ··· + z − pk (z − pk )2 (z − pk )m The coefficients {Aik } can be evaluated through differentiation as illustrated in Example 4.7 for m = 2.

192

The z -Transform and Its Application to the Analysis of LTI Systems

Now that we have performed the partial-fraction expansion, we are ready to take the final step in the inversion of X(z). First, let us consider the case in which X(z) contains distinct poles. From the partial-fraction expansion (4.15), it easily follows that X(z) = A1

1 1 1 + A2 + · · · + AN 1 − p1 z−1 1 − p2 z−1 1 − pN z−1

(4.27)

The inverse z-transform, x(n) = Z −1 {X(z)}, can be obtained by inverting each term in (4.27) and taking the corresponding linear combination. From Table 3 it follows that these terms can be inverted using the formula

Z −1



1 1 − pk z−1



 (pk )n u(n),  

if ROC: |z| > |pk | (causal signals) = n −(p ) u(−n − 1), if ROC: |z| < |pk |  k  (anticausal signals)

(4.28)

If the signal x(n) is causal, the ROC is |z| > pmax , where pmax = max{|p1 |, |p2 |, . . . , |pN |}. In this case all terms in (4.27) result in causal signal components and the signal x(n) is given by n )u(n) x(n) = (A1 p1n + A2 p2n + · · · + AN pN

(4.29)

If all poles are real , (4.29) is the desired expression for the signal x(n). Thus a causal signal, having a z-transform that contains real and distinct poles, is a linear combination of real exponential signals. Suppose now that all poles are distinct but some of them are complex. In this case some of the terms in (4.27) result in complex exponential components. However, if the signal x(n) is real, we should be able to reduce these terms into real components. If x(n) is real, the polynomials appearing in X(z) have real coefficients. In this case, as we have seen in Section 3, if p j is a pole, its complex conjugate pj∗ is also a pole. As was demonstrated in Example 4.6, the corresponding coefficients in the partial-fraction expansion are also complex conjugates. Thus the contribution of two complex-conjugate poles is of the form xk (n) = [Ak (pk )n + A∗k (pk∗ )n ]u(n)

(4.30)

These two terms can be combined to form a real signal component. First, we express Aj and pj in polar form (i.e., amplitude and phase) as Ak = |Ak |ej αk

(4.31)

pk = rk ejβk

(4.32)

where αk and βk are the phase components of Ak and pk . Substitution of these relations into (4.30) gives xk (n) = |Ak |rkn [ej (βk n+αk ) + e−j (βk n+αk ) ]u(n)

193

The z -Transform and Its Application to the Analysis of LTI Systems

or, equivalently, xk (n) = 2|Ak |rkn cos(βk n + αk )u(n) Thus we conclude that   A∗k Ak −1 Z + = 2|Ak |rkn cos(βk n + αk )u(n) 1 − pk z−1 1 − pk∗ z−1

(4.33)

(4.34)

if the ROC is |z| > |pk | = rk . From (4.34) we observe that each pair of complex-conjugate poles in the zdomain results in a causal sinusoidal signal component with an exponential envelope. The distance rk of the pole from the origin determines the exponential weighting (growing if rk > 1, decaying if rk < 1, constant if rk = 1). The angle of the poles with respect to the positive real axis provides the frequency of the sinusoidal signal. The zeros, or equivalently the numerator of the rational transform, affect only indirectly the amplitude and the phase of xk (n) through Ak . In the case of multiple poles, either real or complex, the inverse transform of terms of the form A/(z − pk )n is required. In the case of a double pole the following transform pair (see Table 3) is quite useful:   pz−1 −1 Z = npn u(n) (4.35) (1 − pz−1 )2 provided that the ROC is |z| > |p|. The generalization to the case of poles with higher multiplicity is obtained by using multiple differentiation. EXAMPLE 4.8 Determine the inverse z-transform of X(z) =

1 1 − 1.5z−1 + 0.5z−2

if (a) ROC: |z| > 1 (b) ROC: |z| < 0.5 (c) ROC: 0.5 < |z| < 1 Solution. This is the same problem that we treated in Example 4.2. The partial-fraction expansion for X (z) was determined in Example 4.5. The partial-fraction expansion of X(z) yields 2 1 X(z) = − (4.36) −1 1−z 1 − 0.5z−1 To invert X( z) we should apply (4.28) for p 1 = 1 and p2 = 0.5. However, this requires the specification of the corresponding ROC. (a) In the case when the ROC is |z| > 1, the signal x(n) is causal and both terms in (4.36) are causal terms. According to (4.28), we obtain x(n) = 2(1)n u(n) − (0.5)n u(n) = (2 − 0.5n )u(n) which agrees with the result in Example 4.2(a).

194

(4.37)

The z -Transform and Its Application to the Analysis of LTI Systems

(b) When the ROC is |z| < 0.5, the signal x(n) is anticausal. Thus both terms in (4.36) result in anticausal components. From (4.28) we obtain x(n) = [−2 + (0.5)n ]u(−n − 1)

(4.38)

(c) In this case the ROC 0.5 < |z| < 1 is a ring, which implies that the signal x(n) is two-sided. Thus one of the terms corresponds to a causal signal and the other to an anticausal signal. Obviously, the given ROC is the overlapping of the regions |z| > 0.5 and |z| < 1. Hence the pole p2 = 0.5 provides the causal part and the pole p1 = 1 the anticausal. Thus x(n) = −2(1)n u(−n − 1) − (0.5)n u(n)

(4.39)

EXAMPLE 4.9 Determine the causal signal x(n) whose z-transform is given by X(z) = Solution.

1 + z−1 1 − z−1 + 0.5z−2

In Example 4.6 we have obtained the partial-fraction expansion as X(z) =

A2 A1 + −1 1 − p1 z 1 − p2 z−1

where A1 = A∗2 =

1 3 −j 2 2

and 1 1 +j 2 2 Since we have a pair of complex-conjugate poles, we should use (4.34). The polar forms of A1 and p1 are √ 10 −j 71.565 e A1 = 2 p1 = p2∗ =

1 p1 = √ ej π/4 2 Hence x(n) =

πn

√  1 n 10 √ cos − 71.565◦ u(n) 4 2

EXAMPLE 4.10 Determine the causal signal x(n) having the z-transform X(z) = Solution.

1 (1 + z−1 )(1 − z−1 )2

From Example 4.7 we have X(z) =

1 1 z−1 3 1 1 + + −1 −1 41+z 41−z 2 (1 − z−1 )2

195

The z -Transform and Its Application to the Analysis of LTI Systems

By applying the inverse transform relations in (4.28) and (4.35), we obtain  1 1 1 3 3 n x(n) = (−1)n u(n) + u(n) + nu(n) = (−1)n + + u(n) 4 4 2 4 4 2

4.4

Decomposition of Rational z -Transforms

At this point it is appropriate to discuss some additional issues concerning the decomposition of rational z-transforms, which will prove very useful in the implementation of discrete-time systems. Suppose that we have a rational z-transform X(z) expressed as M 

X(z) =

M 

bk z−k

k=0 N 

1+

= b0 ak z

−k

k=1

(1 − zk z−1 )

k=1 N 

(4.40) −1

(1 − pk z )

k=1

where, for simplicity, we have assumed that a0 ≡ 1. If M ≥ N [i.e., X(z) is improper], we convert X(z) to a sum of a polynomial and a proper function X(z) =

M−N 

ck z−k + Xpr (z)

(4.41)

k=0

If the poles of Xpr (z) are distinct, it can be expanded in partial fractions as Xpr (z) = A1

1 1 1 + A2 + · · · + AN −1 −1 1 − p1 z 1 − p2 z 1 − pN z−1

(4.42)

As we have already observed, there may be some complex-conjugate pairs of poles in (4.42). Since we usually deal with real signals, we should avoid complex coefficients in our decomposition. This can be achieved by grouping and combining terms containing complex-conjugate poles, in the following way: A A∗ A − Ap ∗ z−1 + A∗ − A∗ pz−1 + = 1 − pz−1 1 − p ∗ z−1 1 − pz−1 − p ∗ z−1 + pp ∗ z−2 b0 + b1 z−1 = 1 + a1 z−1 + a2 z−2 where

b0 = 2 Re(A), b1 = 2 Re(Ap∗),

(4.43)

a1 = −2 Re(p) a2 = |p|2

(4.44)

are the desired coefficients. Obviously, any rational transform of the form (4.43) with coefficients given by (4.44), which is the case when a 12 − 4a2 < 0, can be inverted using (4.34). By combining (4.41), (4.42), and (4.43) we obtain a

196

The z -Transform and Its Application to the Analysis of LTI Systems

partial-fraction expansion for the z-transform with distinct poles that contains real coefficients. The general result is X(z) =

M−N 

ck z

k=0

−k

+

K1  k=1

2  bk b0k + b1k z−1 + 1 + ak z−1 1 + a1k z−1 + a2k z−2

K

(4.45)

k=1

where K1 + 2K2 = N . Obviously, if M = N , the first term is just a constant, and when M < N , this term vanishes. When there are also multiple poles, some additional higher-order terms should be included in (4.45). Analternative form is obtained by expressing X(z) as a product of simple terms as in (4.40) . However, the complex-conjugate poles and zeros should be combined to avoid complex coefficients in the decomposition. Such combinations result in second-order rational terms of the following form: (1 − zk z−1 )(1 − zk∗ z−1 ) 1 + b1k z−1 + b2k z−2 ∗ −1 = −1 (1 − pk z )(1 − pk z ) 1 + a1k z−1 + a2k z−2 where

b1k = −2 Re(zk ), b2k = |zk |2 ,

(4.46)

a1k = −2 Re(pk ) a2k = |pk |2

(4.47)

Assuming for simplicity that M = N , we see that X(z) can be decomposed in the following way: K1 K2  1 + bk z−1  1 + b1k z−1 + b2k z−2 X(z) = b0 1 + ak z−1 1 + a1k z−1 + a2k z−2 k=1

(4.48)

k=1

where N = K1 + 2K2 .

5

Analysis of Linear Time-Invariant Systems in the z -Domain In Section 3.3 we introduced the system function of a linear time-invariant system and related it to the unit sample response and to the difference equation description of systems. In this section we describe the use of the system function in the determination of the response o f the system to some excitation signal. In Section 6.3, we extend this method of analysis to nonrelaxed systems. Our attention is focused on the important class of pole–zero systems represented by linear constant-coefficient difference equations with arbitrary initial conditions. We also consider the topic of stability of linear time-invariant systems and describe a test for determining the stability of a system based on the coefficients of the denominator polynomial in the system function. Finally, we provide a detailed analysis of second-order systems, which form the basic building blocks in the realization of higher-order systems.

197

The z -Transform and Its Application to the Analysis of LTI Systems

5.1

Response of Systems with Rational System Functions

Let us consider a pole–zero system described by the general linear constant-coefficient difference equation in (3.7) and the corresponding system function i n (3 .8). We represent H (z) as a ratio of two polynomials B(z)/A(z), where B(z) is the numerator polynomial that contains the zeros of H (z), and A(z) is the denominator polynomial that determines the poles of H (z). Furthermore, let us assume that the input signal x(n) has a rational z-transform X(z) of the form X(z) =

N (z) Q(z)

(5.1)

This assumption is not overly restrictive, since, as indicated previously, most signals of practical interest have rational z-transforms. If the system is initially relaxed, that is, the initial conditions for the difference equation are zero, y(−1) = y(−2) = · · · = y(−N ) = 0, the z-transform of the output of the system has the form Y (z) = H (z)X(z) =

B(z)N (z) A(z)Q(z)

(5.2)

Now suppose that the system contains simple poles p1 , p2 , . . . , pN and the z-transform of the input signal contains poles q1 , q2 , . . . , qL , where pk = qm for all k = 1, 2, . . . , N and m = 1, 2, . . . , L. In addition, we assume that the zeros of the numerator polynomials B(z) and N (z) do not coincide with the poles {pk } and {qk }, so that there is no pole–zero cancellation. Then a partial-fraction expansion of Y (z) yields Y (z) =

N  k=1

 Ak Qk + −1 1 − pk z 1 − qk z−1 L

(5.3)

k=1

The inverse transform of Y (z) yields the output signal from the system in the form y(n) =

N  k=1

Ak (pk )n u(n) +

L 

Qk (qk )n u(n)

(5.4)

k=1

We observe that the output sequence y(n) can be subdivided into two parts. The first part is a function of the poles {pk } of the system and is called the natural response of the system. The influence of the input signal on this part of the response is through the scale factors {Ak }. The second part of the response is a function of the poles {qk } of the input signal and is called the forced response of the system. The influence of the system on this response is exerted through the scale factors {Qk }. We should emphasize that the scale factors {Ak } and {Qk } are functions of both sets of poles {pk } and {qk }. For example, if X(z) = 0 so that the input is zero, then Y (z) = 0, and consequently, the output is zero. Clearly, then, the natural response of the system is zero. This implies that the natural response of the system is different from the zero-input response.

198

The z -Transform and Its Application to the Analysis of LTI Systems

When X(z) and H (z) have one or more poles in common or when X(z) and/or H (z) contain multiple-order poles, then Y (z) will have multiple-order poles. Consequently, the partial-fraction expansion of Y (z) will contain factors of the form 1/(1 − pl z−1 )k , k = 1, 2, . . . , m, where m is the pole order. The inversion of these factors will produce terms of the form nk−1 pln in the output y(n) of the system, as indicated in Section 4.3.

5.2

Transient and Steady-State Responses

As we have seen from our previous discussion, the zero-state response of a system to a given input can be separated into two components, the natural response and the forced response. The natural response of a causal system has the form ynr (n) =

N 

Ak (pk )n u(n)

(5.5)

k=1

where {pk }, k = 1, 2, . . . , N are the poles of the system and {Ak } are scale factors that depend on the initial conditions and on the characteristics of the input sequence. If |pk | < 1 for all k, then, ynr (n) decays to zero as n approaches infinity. In such a case we refer to the natural response of the system as the transient response. The rate at which ynr (n) decays toward zero depends on the magnitude of the pole positions. If all the poles have small magnitudes, the decay is very rapid. On the other hand, if one or more poles are located near the unit circle, the corresponding terms in ynr (n) will decay slowly toward zero and the transient will persist for a relatively long time. The forced response of the system has the form yfr (n) =

L 

Qk (qk )n u(n)

(5.6)

k=1

where {qk }, k = 1, 2, . . . , L are the poles in the forcing function and {Qk } are scale factors that depend on the input sequence and on the characteristics of the system. If all the poles of the input signal fall inside the unit circle, yfr (n) will decay toward zero as n approaches infinity, just as in the case of the natural response. This should not be surprising since the input signal is also a transient signal. On the other hand, when the causal input signal is a sinusoid, the poles fall on the unit circle and consequently, the forced response is also a sinusoid that persists for all n ≥ 0. In this case, the forced response is called the steady-state response of the system. Thus, for the system to sustain a steady-state output for n ≥ 0, the input signal must persist for all n ≥ 0. The following example illustrates the presence of the steady-state response. EXAMPLE 5.1 Determine the transient and steady-state responses of the system characterized by the difference equation y(n) = 0.5y(n − 1) + x(n) when the input signal is x(n) = 10 cos(πn/4)u(n). The system is initially at rest (i.e., it is relaxed).

199

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

The system function for this system is H (z) =

1 1 − 0.5z−1

and therefore the system has a pole at z = 0.5. The z-transform of the input signal is (from Table 3) √ 10(1 − (1/ 2)z−1 ) X(z) = √ 1 − 2z−1 + z−2 Consequently, Y (z) = H (z)X(z) =

√ 10(1 − (1/ 2)z−1 ) (1 − 0.5z−1 )(1 − ej π/4 z−1 )(1 − e−j π/4 z−1 )

=

6.78e−j 28.7 6.78ej 28.7 6.3 + + −1 j π/4 −1 1 − 0.5z 1−e z 1 − e−j π/4 z−1





The natural or transient response is ynr (n) = 6.3(0.5)n u(n) and the forced or steady-state response is yfr (n) = [6.78e−j 28.7 (ej π n/4 ) + 6.78ej 28.7 e−j π n/4 ]u(n)

π n − 28.7◦ u(n) = 13.56 cos 4 Thus we see that the steady-state response persists for all n ≥ 0, just as the input signal persists for all n ≥ 0.

5.3

Causality and Stability

As defined previously, a causal linear time-invariant system is one whose unit sample response h(n) satisfies the condition h(n) = 0,

n r < 1. Since the ROC cannot contain any poles of H (z), it follows that a causal linear time-invariant system is BIBO stable if and only if all the poles of H (z) are inside the unit circle. EXAMPLE 5.2 A linear time-invariant system is characterized by the system function H (z) = =

3 − 4z−1 1 − 3.5z−1 + 1.5z−2 1 1−

1 −1 z 2

+

2 1 − 3z−1

Specify the ROC of H (z) and determine h(n) for the following conditions: (a) The system is stable. (b) The system is causal. (c) The system is anticausal.

201

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

The system has poles at z =

1 2

and z = 3.

(a) Since the system is stable, its ROC must include the unit circle and hence it is Consequently, h(n) is noncausal and is given as

1 2

< |z| < 3.

1 h(n) = ( )n u(n) − 2(3)n u(−n − 1) 2 (b) Since the system is causal, its ROC is |z| > 3. In this case 1 h(n) = ( )n u(n) + 2(3)n u(n) 2 This system is unstable. (c) If the system is anticausal, its ROC is |z| < 0.5. Hence 1 h(n) = −[( )n + 2(3)n ]u(−n − 1) 2 In this case the system is unstable.

5.4

Pole–Zero Cancellations

When a z-transform has a pole that is at the same location as a zero, the pole is canceled by the zero and, consequently, the term containing that pole in the inverse z-transform vanishes. Such pole–zero cancellations are very important in the analysis of pole–zero systems. Pole–zero cancellations can occur either in the system function itself or in the product of the system function with the z-transform of the input signal. In the first case we say that the order of the system is reduced by one. In the latter case we say that the pole of the system is suppressed by the zero in the input signal, or vice versa. Thus, by properly selecting the position of the zeros of the input signal, it is possible to suppress one or more system modes (pole factors) in the response of the system. Similarly, by proper selection of the zeros of the system function, it is possible to suppress one or more modes of the input signal from the response of the system. When the zero is located very near the pole but not exactly at the same location, the term in the response has a very small amplitude. For example, nonexact pole– zero cancellations can occur in practice as a result of insufficiant numerical precision used in representing the coefficients of the system. Consequently, one should not attempt to stabilize an inherently unstable system by placing a zero in the input signal at the location of the pole. EXAMPLE 5.3 Determine the unit sample response of the system characterized by the difference equation y(n) = 2.5y(n − 1) − y(n − 2) + x(n) − 5x(n − 1) + 6x(n − 2)

202

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

The system function is H (z) = =

1 − 5z−1 + 6z−2 1 − 2.5z−1 + z−2 1 − 5z−1 + 6z−2 (1 − 21 z−1 )(1 − 2z−1 )

This system has poles at p1 = 2 and p1 = 21 . Consequently, at first glance it appears that the unit sample response is Y (z) = H (z)X(z) =  =z By evaluating the constants at z =

A z−

1 2

1 2

1 − 5z−1 + 6z−2

(1 − 21 z−1 )(1 − 2z−1 )  B + z−2

and z = 2, we find that A=

5 , 2

B=0

The fact that B = 0 indicates that there exists a zero at z = 2 which cancels the pole at z = 2. In fact, the zeros occur at z = 2 and z = 3. Consequently, H (z) reduces to H (z) =

1 − 3z−1 1−

=1−

1 −1 z 2

=

z−3 z−

1 2

2.5z−1 1 − 21 z−1

and therefore 1 h(n) = δ(n) − 2.5( )n−1 u(n − 1) 2 The reduced-order system obtained by canceling the common pole and zero is characterized by the difference equation y(n) =

1 y(n − 1) + x(n) − 3x(n − 1) 2

Although the original system is also BIBO stable due to the pole–zero cancellation, in a practical implementation of this second-order system, we may encounter an instability due to imperfect cancellation of the pole and the zero.

EXAMPLE 5.4 Determine the response of the system y(n) =

1 5 y(n − 1) − y(n − 2) + x(n) 6 6

to the input signal x(n) = δ(n) − 13 δ(n − 1).

203

The z -Transform and Its Application to the Analysis of LTI Systems

Solution.

The system function is H (z) =

1 1−

=  This system has two poles, one at z = signal is

1 2

5 −1 z 6

1−

+ 16 z−2

1 −1 z 2

1 

1 − 13 z−1



and the other at z = 13 . The z-transform of the input

1 X(z) = 1 − z−1 3 In this case the input signal contains a zero at z = Consequently,

1 3

which cancels the pole at z =

1 . 3

Y (z) = H (z)X(z) Y (z) =

1 1 − 21 z−1

and hence the response of the system is 1 y(n) = ( )n u(n) 2 Clearly, the mode ( 13 )n is suppressed from the output as a result of the pole–zero cancellation.

5.5

Multiple-Order Poles and Stability

As we have observed, a necessary and sufficient condition for a causal linear timeinvariant system to be BIBO stable is that all its poles lie inside the unit circle. The input signal is bounded if its z-transform contains poles {qk }, k = 1, 2, . . . , L, which satisfy the condition |qk | ≤ 1 for all k. We note that the forced response of the system, given in (5.6), is also bounded, even when the input signal contains one or more distinct poles on the unit circle. In view of the fact that a bounded input signal may have poles on the unit circle, it might appear that a stable system may also have poles on the unit circle. This is not the case, however, since such a system produces an unbounded response when excited by an input signal that also has a pole at the same position on the unit circle. The following example illustrates this point. EXAMPLE 5.5 Determine the step response of the causal system described by the difference equation y(n) = y(n − 1) + x(n) Solution.

The system function for the system is H (z) =

204

1 1 − z−1

The z -Transform and Its Application to the Analysis of LTI Systems

We note that the system contains a pole on the unit circle at z = 1. The z-transform of the input signal x(n) = u(n) is 1 X(z) = 1 − z−1 which also contains a pole at z = 1. Hence the output signal has the transform Y (z) = H (z)X(z) =

1 (1 − z−1 )2

which contains a double pole at z = 1. The inverse z-transform of Y (z) is y(n) = (n + 1)u(n) which is a ramp sequence. Thus y(n) is unbounded, even when the input is bounded. Consequently, the system is unstable.

Example 5.5 demonstrates clearly that BIBO stability requires that the system poles be strictly inside the unit circle. If the system poles are all inside the unit circle and the excitation sequence x(n) contains one or more poles that coincide with the poles of the system, the output Y (z) will contain multiple-order poles. As indicated previously, such multiple-order poles result in an output sequence that contains terms of the form Ak nb (pk )n u(n) where 0 ≤ b ≤ m − 1 and m is the order of the pole. If |pk | < 1, these terms decay to zero as n approaches infinity because the exponential factor (pk )n dominates the term nb . Consequently, no bounded input signal can produce an unbounded output signal if the system poles are all inside the unit circle. Finally, we should state that the only useful systems which contain poles on the unit circle are digital oscillators. We call such systems marginally stable.

5.6

Stability of Second-Order Systems

In this section we provide a detailed analysis of a system having two poles. Two-pole systems form the basic building blocks for the realization of higher-order systems. Let us consider a causal two-pole system described by the second-order difference equation y(n) = −a1 y(n − 1) − a2 y(n − 2) + b0 x(n) (5.7) The system function is H (z) =

b0 Y (z) = −1 X(z) 1 + a1 z + a2 z−1

b0 z 2 = 2 z + a1 z + a 2

(5.8)

205

The z -Transform and Its Application to the Analysis of LTI Systems

This system has two zeros at the origin and poles at  a1 p1 , p2 = − ± 2

a12 − 4a2 4

(5.9)

The system is BIBO stable if the poles lie inside the unit circle, that is, if |p1 | < 1 and |p2 | < 1. These conditions can be related to the values of the coefficients a1 and a2 . In particular, the roots of a quadratic equation satisfy the relations a1 = −(p1 + p2 )

(5.10)

a2 = p1 p2

(5.11)

From (5.10) and (5.11) we easily obtain the conditions that a1 and a2 must satisfy for stability. First, a2 must satisfy the condition |a2 | = |p1 p2 | = |p1 ||p2 | < 1

(5.12)

The condition for a1 can be expressed as |a1 | < 1 + a2

(5.13)

Therefore, a two-pole system is stable if and only if the coefficients a1 and a2 satisfy the conditions in (5.12) and (5.13). The stability conditions given in (5.12) and (5.13) define a region in the coefficient plane (a1 , a2 ), which is in the form of a triangle, as shown in Fig. 5.1. The system is stable if and only if the point (a1 , a2 ) lies inside the triangle, which we call the stability triangle. a2 a2 =

Complexconjugate poles

Stability triangle

a21

4 a2 = a1 − 1

1 a2 = 1 Real and equal poles −2

−1

1 −1

a1 2 Real and distinct poles

a2 = −a1 − 1

Figure 5.1 Region of stability (stability triangle) in the

(a1 , a2 ) coefficient plane for a second-order system.

206

The z -Transform and Its Application to the Analysis of LTI Systems

The characteristics of the two-pole system depend on the location of the poles or, equivalently, on the location of the point (a1 , a2 ) in the stability triangle. The poles of the system may be real or complex conjugate, depending on the value of the discriminant  = a12 − 4a2 . The parabola a2 = a12 /4 splits the stability triangle into two regions, as illustrated in Fig. 5.1. The region below the parabola (a12 > 4a2 ) corresponds to real and distinct poles. The points on the parabola (a12 = 4a2 ) result in real and equal (double) poles. Finally, the points above the parabola correspond to complex-conjugate poles. Additional insight into the behavior of the system can be obtained from the unit sample responses for these three cases. Real and distinct poles ( a12 > 4a2 ). Since p1 , p2 are real and p1 = p2 , the system

function can be expressed in the form H (z) = where A1 =

A2 A1 + −1 1 − p1 z 1 − p2 z−1

(5.14)

−b0 p2 p1 − p 2

(5.15)

b0 (p n+1 − p2n+1 )u(n) p1 − p 2 1

(5.16)

b0 p1 , p1 − p 2

A2 =

Consequently, the unit sample response is h(n) =

Therefore, the unit sample response is the difference of two decaying exponential sequences. Figure 5.2 illustrates a typical graph for h(n) when the poles are distinct. Real and equal poles (a12 = 4a2 ) .

In this case p1 = p2 = p = −a1 /2. The system

function is

H (z) =

b0 (1 − pz−1 )2

(5.17)

h(n) 2.0

1.5

1.0

0.5

0

50

n

Figure 5.2 Plot of h(n) given by (5.16) with p 1 = 0.5, p2 = 0.75; h(n) = [1/(p1 − p2 )](p1n+1 − p2n+1 )u(n).

207

The z -Transform and Its Application to the Analysis of LTI Systems

and hence the unit sample response of the system is h(n) = b0 (n + 1)p n u(n)

(5.18)

We observe that h(n) is the product of a ramp sequence and a real decaying exponential sequence. The graph of h(n) is shown in Fig. 5.3. Complex-conjugate poles ( a12 < 4a2 ).

Since the poles are complex conjugate, the system function can be factored and expressed as H (z) =

A∗ A + 1 − pz−1 1 − p ∗ z−1

(5.19)

A A∗ = + 1 − rej ω0 z−1 1 − re−j ω0 z−1

where p = rej ω and 0 < ω0 < π . Note that when the poles are complex conjugates, the parameters a1 and a2 are related to r and ω0 according to a1 = −2r cos ω0 (5.20)

a2 = r 2

The constant A in the partial-fraction expansion of H (z) is easily shown to be A=

b0 p b0 rej ω0 = p − p∗ r(ej ω0 − e−j ω0 )

(5.21)

b 0 e j ω0 = j 2 sin ω0 h(n) 2.0

1.5

1.0

0.5

0

Figure 5.3 Plot of h(n) given by (5.18) with p = n

1)p u(n).

208

50 3 ; 4

h(n) = (n +

n

The z -Transform and Its Application to the Analysis of LTI Systems

h(n) 1.2 1.0 0.8 0.6 0.4 0.2 0

50

−0.2

n

−0.4 −0.6 −0.8 −1.0 −1.2

Figure 5.4 Plot of h(n) given by (5.22) with b 0 = 1, ω0 = π/4,

r = 0.9; h(n) = [b0 r n /(sin ω0 )] sin[(n + 1)ω0 ]u(n).

Consequently, the unit sample response of a system with complex-conjugate poles is h(n) =

b0 r n ej (n+1)ω0 − e−j (n+1)ω0 u(n) sin ω0 2j

b0 r n sin(n + 1)ω0 u(n) = sin ω0

(5.22)

In this case h(n) has an oscillatory behavior with an exponentially decaying envelope when r < 1. The angle ω0 of the poles determines the frequency of oscillation and the distance r of the poles from the origin determines the rate of decay. When r is close to unity, the decay is slow. When r is close to the origin, the decay is fast. A typical graph of h(n) is illustrated in Fig. 5.4.

6

The One-sided z -Transform The two-sided z-transform requires that the corresponding signals be specified for the entire time range −∞ < n < ∞. This requirement prevents its use for a very useful family of practical problems, namely the evaluation of the output of nonrelaxed systems. As we recall, these systems are described by difference equations with nonzero initial conditions. Since the input is applied at a finite time, say n0 , both input and output signals are specified for n ≥ n0 , but by no means are zero for n < n0 . Thus the two-sided z-transform cannot be used. In this section we develop the one-sided z-transform which can be used to solve difference equations with initial conditions.

209

The z -Transform and Its Application to the Analysis of LTI Systems

6.1

Definition and Properties

The one-sided or unilateral z-transform of a signal x(n) is defined by X + (z) ≡

∞ 

x(n)z−n

(6.1)

n=0

We also use the notations Z + {x(n)} and z+

x(n) ←→ X + (z) The one-sided z-transform differs from the two-sided transform in the lower limit of the summation, which is always zero, whether or not the signal x(n) is zero for n < 0 (i.e., causal). Due to this choice of lower limit, the one-sided z-transform has the following characteristics: 1. It does not contain information about the signal x(n) for negative values of time (i.e., for n < 0). 2. It is unique only for causal signals, because only these signals are zero for n < 0. 3. The one-sided z-transform X+ (z) of x(n) is identical to the two-sided z-transform of the signal x(n)u(n). Since x(n)u(n) is causal, the ROC of its transform, and hence the ROC of X + (z), is always the exterior of a circle. Thus when we deal with one-sided z-transforms, it is not necessary to refer to their ROC. EXAMPLE 6.1 Determine the one-sided z-transform of the signals in Example 1.1. Solution.

From the definition (6.1), we obtain z+

x1 (n) = {1, 2, 5, 7, 0, 1} ←→ X1+ (z) = 1 + 2z−1 + 5z−2 + 7z−3 + z−5 ↑

z+

x2 (n) = {1, 2, 5, 7, 0, 1} ←→ X2+ (z) = 5 + 7z−1 + z−3 ↑

z+

x3 (n) = {0, 0, 1, 2, 5, 7, 0, 1} ←→ X3+ (z) = z−2 + 2z−3 + 5z−4 + 7z−5 + z−7 ↑

z+

x4 (n) = {2, 4, 5, 7, 0, 1} ←→ X4+ (z) = 5 + 7z−1 + z−3 ↑ z+

x5 (n) = δ(n) ←→ X5+ (z) = 1 z+

x6 (n) = δ(n − k),

k > 0 ←→ X6+ (z) = z−k

x7 (n) = δ(n + k),

k > 0 ←→ X7+ (z) = 0

z+

Note that for a noncausal signal, the one-sided z-transform is not unique. Indeed, X2+ (z) = X4+ (z) but x2 (n) = x4 (n). Also for anticausal signals, X+ (z) is always zero.

210

The z -Transform and Its Application to the Analysis of LTI Systems

Almost all properties we have studied for the two-sided z-transform carry over to the one-sided z-transform with the exception of the shifting property. Shfiting Property Case 1: Time delay

If z+

x(n) ←→ X + (z) then z+

x(n − k) ←→ z−k [X + (z) +

k 

x(−n)zn ],

k>0

(6.2)

n=1

In case x(n) is causal, then z+

x(n − k) ←→ z−k X + (z)

Proof

(6.3)

From the definition (6.1) we have

−1 ∞   x(l)z−l + x(l)z−l Z + {x(n − k)} = z−k l=−k

 = z−k 

−k 

l=0



x(l)z−l + X + (z)

l=−1

By changing the index from l to n = −l , the result in (6.2) is easily obtained. EXAMPLE 6.2 Determine the one-sided z-transform of the signals (a) x(n) = a n u(n) (b) x1 (n) = x(n − 2) where x(n) = a n Solution. (a) From (6.1) we easily obtain X+ (z) =

1 1 − az−1

(b) We will apply the shifting property for k = 2. Indeed, we have Z + {x(n − 2)} = z−2 [X+ (z) + x(−1)z + x(−2)z2 ] = z−2 X+ (z) + x(−1)z−1 + x(−2) Since x(−1) = a −1 , x(−2) = a −2 , we obtain X1+ (z) =

z−2 + a −1 z−1 + a −2 1 − az−1

211

The z -Transform and Its Application to the Analysis of LTI Systems

The meaning of the shifting property can be intuitively explained if we write (6.2) as follows: Z + {x(n − k)} = [x(−k) + x(−k + 1)z−1 + · · · + x(−1)z−k+1 ] + z−k X + (z),

(6.4)

k>0

To obtain x(n − k)(k > 0) from x(n), we should shift x(n) by k samples to the right. Then k “new” samples, x(−k), x(−k + 1), . . . , x(−1), enter the positive time axis with x(−k) located at time zero. The first term in (6.4) stands for the z-transform of these samples. The “old” samples of x(n − k) are the same as those of x(n) simply shifted by k samples to the right. Their z-transform is obviously z−k X + (z), which is the second term in (6.4). Case 2: Time advance

If

z+

x(n) ←→ X + (z) then z+



x(n + k) ←→ z

k

+

X (z) −

k−1 

x(n)z

−n

k>0

,

(6.5)

n=0

Proof

From (6.1) we have Z + {x(n + k)} =

∞ 

x(n + k)z−n = zk

∞ 

x(l)z−l

l=k

n=0

where we have changed the index of summation from n to l = n + k. Now, from (6.1) we obtain +

X (z) =

∞ 

x(l)z

−l

=

l=0

k−1 

x(l)z

−l

l=0

+

∞ 

x(l)z−l

l=k

By combining the last two relations, we easily obtain (6.5). EXAMPLE 6.3 With x(n), as given in Example 6.2, determine the one-sided z-transform of the signal x2 (n) = x(n + 2) Solution.

We will apply the shifting theorem for k = 2. From (6.5), with k = 2, we obtain Z + {x(n + 2)} = z2 X+ (z) − x(0)z2 − x(1)z

But x(0) = 1, x(1) = a , and X+ (z) = 1/(1 − az−1 ). Thus Z + {x(n + 2)} =

212

z2 − z2 − az 1 − az−1

The z -Transform and Its Application to the Analysis of LTI Systems

The case of a time advance can be intuitively explained as follows. To obtain x(n+k), k > 0, we should shift x(n) by k samples to the left. As a result, the samples x(0), x(1), . . . , x(k − 1) “leave” the positive time axis. Thus we first remove their contribution to the X+ (z), and then multiply what remains by zk to compensate for the shifting of the signal by k samples. The importance of the shifting property lies in its application to the solution of difference equations with constant coefficients and nonzero initial conditions. This makes the one-sided z-transform a very useful tool for the analysis of recursive linear time-invariant discrete-time systems. An important theorem useful in the analysis of signals and systems is the final value theorem. Final Value Theorem. If

z+

x(n) ←→ X + (z) then

lim x(n) = lim (z − 1)X + (z)

n→∞

(6.6)

z→1

The limit in (6.6) exists if the ROC of (z − 1)X+ (z) includes the unit circle. The proof of this theorem is left as an exercise for the reader. This theorem is useful when we are interested in the asymptotic behavior of a signal x(n) and we know its z-transform, but not the signal itself. In such cases, especially if it is complicated to invert X+ (z), we can use the final value theorem to determine the limit of x(n) as n goes to infinity. EXAMPLE 6.4 The impulse response of a relaxed linear time-invariant system is h(n) = α n u(n), |α| < 1. Determine the value of the step response of the system as n → ∞. Solution.

The step response of the system is y(n) = h(n) ∗ x(n)

where x(n) = u(n) Obviously, if we excite a causal system with a causal input the output will be causal. Since h(n), x(n), y(n) are causal signals, the one-sided and two-sided z-transforms are identical. From the convolution property (2.17) we know that the z-transforms of h(n) and x(n) must be multiplied to yield the z-transform of the output. Thus Y (z) =

1 z2 1 = , −1 −1 1 − αz 1 − z (z − 1)(z − α)

ROC: |z| > |α|

Now

z2 , ROC: |z| < |α| z−α Since |α| < 1, the ROC of (z − 1)Y (z) includes the unit circle. Consequently, we can apply (6.6) and obtain z2 1 = lim y(n) = lim n→∞ z→1 z − α 1−α (z − 1)Y (z) =

213

The z -Transform and Its Application to the Analysis of LTI Systems

6.2

Solution of Difference Equations

The one-sided z-transform is a very efficient tool for the solution of difference equations with nonzero initial conditions. It achieves that by reducing the difference equation relating the two time-domain signals to an equivalent algebraic equation relating their one-sided z-transforms. This equation can be easily solved to obtain the transform of the desired signal. The signal in the time domain is obtained by inverting the resulting z-transform. We will illustrate this approach with two examples. EXAMPLE 6.5 The well-known Fibonacci sequence of integer numbers is obtained by computing each term as the sum of the two previous ones. The first few terms of the sequence are 1, 1, 2, 3, 5, 8, . . . Determine a closed-form expression for the nth term of the Fibonacci sequence. Solution. Let y(n) be the nth term of the Fibonacci sequence. Clearly, y(n) satisfies the difference equation y(n) = y(n − 1) + y(n − 2) (6.7) with initial conditions y(0) = y(−1) + y(−2) = 1

(6.8a)

y(1) = y(0) + y(−1) = 1

(6.8b)

From (6.8b) we have y(−1) = 0. Then (6.8a) gives y(−2) = 1. Thus we have to determine y(n), n ≥ 0, which satisfies (6.7), with initial conditions y(−1) = 0 and y(−2) = 1. By taking the one-sided z -transform of (6.7) and using the shifting property (6.2), we obtain Y + (z) = [z−1 Y + (z) + y(−1)] + [z−2 Y + (z) + y(−2) + y(−1)z−1 ] or Y + (z) =

1 1−

z−1



z2

=

z2

z2 −z−1

(6.9)

where we have used the fact that y(−1) = 0 and y(−2) = 1. We can invert Y + (z) by the partial-fraction expansion method. The poles of Y + (z) are √ √ 1+ 5 1− 5 p1 = , p2 = 2 2 √ √ and the corresponding coefficients are A1 = p1 / 5 and A2 = −p2 / 5. Therefore,

√  √  √ n √ n 1+ 5 1+ 5 1− 5 1− 5 y(n) = − √ u(n) √ 2 2 2 5 2 5 or, equivalently, 1 y(n) = √ 5

214

 n+1  √ n+1 √ n+1 1 1+ 5 − 1− 5 u(n) 2

(6.10)

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 6.6 Determine the step response of the system y(n) = αy(n − 1) + x(n),

−1 < α < 1

(6.11)

when the initial condition is y(−1) = 1. Solution.

By taking the one-sided z-transform of both sides of (6.11), we obtain Y + (z) = α[z−1 Y + (z) + y(−1)] + X + (z)

Upon substitution for y(−1) and X+ (z) and solving for Y + (z), we obtain the result Y + (z) =

α 1 + 1 − αz−1 (1 − αz−1 )(1 − z−1 )

(6.12)

By performing a partial-fraction expansion and inverse transforming the result, we have y(n) = α n+1 u(n) +

1 − α n+1 u(n) 1−α

(6.13)

1 (1 − α n+2 )u(n) = 1−α

6.3

Response of Pole–Zero Systems with Nonzero Initial Conditions

Suppose that the signal x(n) is applied to the pole–zero system at n = 0. Thus the signal x(n) is assumed to be causal. The effects of all previous input signals to the system are reflected in the initial conditions y(−1), y(−2), . . . , y(−N ). Since the input x(n) is causal and since we are interested in determining the output y(n) for n ≥ 0, we can use the one-sided z-transform, which allows us to deal with the initial conditions. Thus the one-sided z-transform of (7) becomes +

Y (z) = −

N 

ak z

−k

+

Y (z) +

k=1

k 

y(−n)z

n

+

M 

bk z−k X + (z)

(6.14)

k=0

n=1

Since x(n) is causal, we can set X + (z) = X(z). In any case (6.14) may be expressed as k N M    ak z−k y(−n)zn bk z−k Y + (z) =

k=0 N 

1+

X(z) − ak z−k

k=1

= H (z)X(z) +

k=1

n=1

1+

N 

ak z−k

(6.15)

k=1

N0 (z) A(z)

215

The z -Transform and Its Application to the Analysis of LTI Systems

where N0 (z) = −

N 

ak z

k=1

−k

k 

y(−n)zn

(6.16)

n=1

From (6.15) it is apparent that the output of the system with nonzero initial conditions can be subdivided into two parts. The first is the zero-state response of the system, defined in the z-domain as Yzs (z) = H (z)X(z)

(6.17)

The second component corresponds to the output resulting from the nonzero initial conditions. This output is the zero-input response of the system, which is defined in the z-domain as N0 (z) Yzi+ (z) = (6.18) A(z) Hence the total response is the sum of these two output components, which can be expressed in the time domain by determining the inverse z-transforms of Yzs (z) and Yzi (z) separately, and then adding the results. Thus y(n) = yzs (n) + yzi (n)

(6.19)

Since the denominator of Yzi+ (z), is A(z), its poles are p1 , p2 , . . . , pN . Consequently, the zero-input response has the form

yzi (n) =

N 

Dk (pk )n u(n)

(6.20)

k=1

This can be added to (6.4) and the terms involving the poles {p }k can be combined to yield the total response in the form

y(n) =

N  k=1

Ak (pk )n u(n) +

L 

Qk (qk )n u(n)

(6.21)

k=1

where, by definition, Ak = Ak + Dk

(6.22)

This development indicates clearly that the effect of the initial conditions is to alter the natural response of the system through modification of the scale factors {Ak }. There are no new poles introduced by the nonzero initial conditions. Furthermore, there is no effect on the forced response of the system. These important points are reinforced in the following example.

216

The z -Transform and Its Application to the Analysis of LTI Systems

EXAMPLE 6.7 Determine the unit step response of the system described by the difference equation y(n) = 0.9y(n − 1) − 0.81y(n − 2) + x(n) under the following initial conditions y(−1) = y(−2) = 1. Solution.

The system function is H (z) =

1 1 − 0.9z−1 + 0.81z−2

This system has two complex-conjugate poles at p1 = 0.9ej π/3 ,

p2 = 0.9e−j π/3

The z-transform of the unit step sequence is X(z) =

1 1 − z−1

Therefore, Yzs (z) = =

1 (1 − 0.9ej π/3 z−1 )(1 − 0.9e−j π/3 z−1 )(1 − z−1 ) 0.0496 + j 0.542 1.099 0.0496 − j 0.542 + + 1 − 0.9ej π/3 z−1 1 − 0.9e−j π/3 z−1 1 − z−1

and hence the zero-state response is  π

n − 5.2◦ u(n) yzs (n) = 1.099 + 1.088(0.9)n cos 3 For the initial conditions y(−1) = y(−2) = 1, the additional component in the z-transform is Yzi (z) = =

N0 (z) 0.09 − 0.81z−1 = A(z) 1 − 0.9z−1 + 0.81z−2 0.045 + j 0.4936 0.045 − j 0.4936 + 1 − 0.9ej π/3 z−1 1 − 0.9e−j π/3 z−1

Consequently, the zero-input response is yzi (n) = 0.988(0.9)n cos

π 3

n + 87◦ u(n)

In this case the total response has the z-transform Y (z) = Yzs (z) + Yzi (z) =

0.568 + j 0.445 0.568 − j 0.445 1.099 + + 1 − z−1 1 − 0.9ej π/3 z−1 1 − 0.9e−j π/3 z−1

The inverse transform yields the total response in the form π

n + 38◦ u(n) y(n) = 1.099u(n) + 1.44(0.9)n cos 3

217

The z -Transform and Its Application to the Analysis of LTI Systems

7

Summary and References The z-transform plays the same role in discrete-time signals and systems as the Laplace transform does in continuous-time signals and systems. In this chapter we derived the important properties of the z-transform, which are extremely useful in the analysis of discrete-time systems. Of particular importance is the convolution property, which transforms the convolution of two sequences into a product of their z-transforms. In the context of LTI systems, the convolution property results in the product of the z-transform X(z) of the input signal with the system function H (z), where the latter is the z-transform of the unit sample response of the system. This relationship allows us to determine the output of an LTI system in response to an input with transform X(z) by computing the product Y (z) = H (z)X(z) and then determining the inverse z-transform of Y (z) to obtain the output sequence y(n). We observed that many signals of practical interest have rational z-transforms. Moreover, LTI systems characterized by constant-coefficient linear difference equations also possess rational system functions. Consequently, in determining the inverse z-transform, we naturally emphasized the inversion of rational transforms. For such transforms, the partial-fraction expansion method is relatively easy to apply, in conjunction with the ROC, to determine the corresponding sequence in the time domain. We considered the characterization of LTI systems in the z-transform domain. In particular, we related the pole–zero locations of a system to its time-domain characteristics and restated the requirements for stability and causality of LTI systems in terms of the pole locations. We demonstrated that a causal system has a system function H (z) with a ROC |z| > r1 , where 0 < r1 ≤ ∞. In a stable and causal system, the poles of H (z) lie inside the unit circle. On the other hand, if the system is noncausal, the condition for stability requires that the unit circle be contained in the ROC of H (z). Hence a noncausal stable LTI system has a system function with poles both inside and outside the unit circle with an annular ROC that includes the unit circle. Finally, the one-sided z-transform was introduced to solve for the response of causal systems excited by causal input signals with nonzero initial conditions.

Problems 1 Determine the z-transform of the following signals. (a) x(n) = {3, 0, 0, 0, 0, 6, 1, −4} ↑  1 n (b) x(n) = ( 2 ) , n ≥ 5 0, n≤4 2 Determine the z-transforms of the following signals and sketch the corresponding pole–zero patterns. (a) x(n) = (1 + n)u(n) (b) x(n) = (a n + a −n )u(n), a real (c) x(n) = (−1)n 2−n u(n)

218

The z -Transform and Its Application to the Analysis of LTI Systems

(d) (e) (f) (g)

x(n) = (na n sin ω0 n)u(n) x(n) = (na n cos ω0 n)u(n) x(n) = Ar n cos(ω0 n + φ)u(n), 0 < r < 1 x(n) = 21 (n2 + n)( 13 )n−1 u(n − 1)

(h) x(n) = ( 21 )n [u(n) − u(n − 10)] 3 Determine the z-transforms and sketch the ROC of the following signals.  1 n (3) , n≥0 (a) x1 (n) = ( 21 )−n , n < 0  1 n n (b) x2 (n) = ( 3 ) − 2 , n ≥ 0 0, n

1 2

20 (a) Draw the pole–zero pattern for the signal x1 (n) = (r n sin ω0 n)u(n),

0 |1/a|

1 − 41 z−1

, 1 − 16 z−1 − 16 z−2

|z| >

1 2

57 Let x(n) be a sequence with z-transform X(z) =

1 − a2 , (1 − az)(1 − az−1 )

ROC: a > |z| > 1/a

with 0 < a < 1. Determine x(n) by using contour integration. 58 The z-transform of a sequence x(n) is given by X(z) =

z20 (z − 21 )(z − 2)5 (z + 25 )2 (z + 3)

Furthermore it is known that X(z) converges for |z| = 1. (a) Determine the ROC of X(z). (b) Determine x(n) at n = −18. (Hint: Use contour integration.)

227

The z -Transform and Its Application to the Analysis of LTI Systems

Answers to Selected Problems 1

(a) X(z) = 3z5 + 6 + z−1 − 4z−2 ROC : 0 < |z| < ∞

2

(a) X(z) = (d) X(z) = (h) X(z) =

4

(a) X(z) =

1 (1−z−1 )2

[az−1 −(az−1 )3 ] sin w0 [1−(2a cos w0 )z−1 +a2 z−2 ]  10 1− 21 z−1 1− 21 z−1 z−1 (1+z−1 )2

(f) X(z) = 1 − z

, |z| < a

1 2

, |z| > 1 + z−4 − z−5 , z = 0

8

(a) Y (z) =

12

(a) x(n) = [4(2)n − 3 − n]u(n)

14

16

(a) x(n) = [2(−1)n − (−2)n ]u(n)  n  √ n 23 (c) x(n) = − 35 √12 cos π4 n + 10 1 2 sin π4 n +  1 n+1  1 n−1 (j) x(n) = − a u(n) + a u(n − 1)

    n n u(n) (a) y(n) = − 43 41 + 13 + 21

19

(d) y(n) = [−2(n + 1) + 2n+1 ]u(n)    n (b) x(n) = − n1 21 u(n − 1)

24 35

38

42

X(z) 1−z−1

(a) x(n) = [0.136(0.28)n + 0.864(−1.78)n ]u(n)  n  n √   n (a) y(n) = 17 13 76 21 cos π3n + 3 7 3 21 sin   (d) y(n) = √102 sin π2n + π4 u(n)  n  n  n (h) y(n) = 4 21 − n 41 − 3 41 u(n)  n   n u(n) (a) h(n) = 2 21 − 41  1 n 1  1 n 8 y(n) = 3 − 2 2 + 3 4 u(n)  n  n 2 1 (d) h(n) = 2 5 − 5 u(n)     25 1 1 n 4 2 n y(n) = 12 + 4 5 − 3 5 u(n)

 n−1 9  2 n−1 +2 5 u(n − 1) (a) h(n) = − 27 15

πn 3

49

(a) h(n) = b0 δ(n) + (b1 − b0 a1 )(−a1 )n−1 u(n − 1)

b0 −b1 0 +b1 (b) y(n) = b1+a + a11+a (−a1 )n u(n) 1 1    n 4 3 1 n u(n) (d) y(n) = 3 − 8 2 + 247 − 21

56

(a) x(n) =

44

(d) x(n) = 58

228

−2

, |z| >

2

 1 n

2 3 10

u(n)  1 n 7  2

ROC: a < |z| < 1/a x(−18) = −32/15309

10

− 13

n

u(n)

17 20

u(n)

(12)n u(n)

Frequency Analysis of Signals

From Chapter 4 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

229

Frequency Analysis of Signals

The Fourier transform is one of several mathematical tools that is useful in the analysis and design of linear time - invariant (LTI) systems. Another is the Fourier series. These signal representations basically involve the decomposition of the signals in terms of sinusoidal (or complex exponential) components. With such a decomposition, a signal is said to be represented in the frequency domain. As we shall demonstrate, most signals of practical interest can be decomposed into a sum of sinusoidal signal components. For the class of periodic signals, such a decomposition is called a Fourier series. For the class of finite energy signals, the decomposition is called the Fourier transform. These decompositions are extremely important in the analysis of LTI systems because the response of an LTI system to a sinusoidal input signal is a sinusoid of the same frequency but of different amplitude and phase. Furthermore, the linearity property of the LTI system implies that a linear sum of sinusoidal components at the input produces a similar linear sum of sinusoidal components at the output, which differ only in the amplitudes and phases from the input sinusoids. This characteristic behavior of LTI systems renders the sinusoidal decomposition of signals very important. Although many other decompositions of signals are possible, only the class of sinusoidal (or complex exponential) signals possess this desirable property in passing through an LTI system. We begin our study of frequency analysis of signals with the representation of continuous-time periodic and aperiodic signals by means of the Fourier series and the Fourier transform, respectively. This is followed by a parallel treatment of discretetime periodic and aperiodic signals. The properties of the Fourier transform are described in detail and a number of time-frequency dualities are presented.

230

Frequency Analysis of Signals

1

Frequency Analysis of Continuous-Time Signals It is well known that a prism can be used to break up white light (sunlight) into the colors of the rainbow (see Fig. 1.1(a)). In a paper submitted in 1672 to the Royal Society, Isaac Newton used the term spectrum to describe the continuous bands of colors produced by this apparatus. To understand this phenomenon, Newton placed another prism upside-down with respect to the first, and showed that the colors blended back into white light, as in Fig. 1.1(b). By inserting a slit between the two prisms and blocking one or more colors from hitting the second prism, he showed that the remixed light is no longer white. Hence the light passing through the first prism is simply analyzed into its component colors without any other change. However, only if we mix again all of these colors do we obtain the original white light. Later, Joseph Fraunhofer (1787–1826), in making measurements of light emitted by the sun and stars, discovered that the spectrum of the observed light consists of distinct color lines. A few years later (mid-1800s) Gustav Kirchhoff and Robert Bunsen found that each chemical element, when heated to incandescence, radiated its own distinct color of light. As a consequence, each chemical element can be identified by its own line spectrum. From physics we know that each color corresponds to a specific frequency of the visible spectrum. Hence the analysis of light into colors is actually a form of frequency analysis. Frequency analysis of a signal involves the resolution of the signal into its frequency (sinusoidal) components. Instead of light, our signal waveforms are basically functions of time. The role of the prism is played by the Fourier analysis tools that we will develop: the Fourier series and the Fourier transform. The recombination of the sinusoidal components to reconstruct the original signal is basically a Fourier synthesis problem. The problem of signal analysis is basically the same for the case of a signal waveform and for the case of the light from heated chemical composiGlass prism Violet Blue Green Yellow Orange Red

Beam of sunlight

Spectrum (a) Glass prism White light

Figure 1.1

(a) Analysis and (b) synthesis of the white light (sunlight) using glass prisms.

Beam of sunlight

(b)

231

Frequency Analysis of Signals

tions. Just as in the case of chemical compositions, different signal waveforms have different spectra. Thus the spectrum provides an “identity” or a signature for the signal in the sense that no other signal has the same spectrum. As we will see, this attribute is related to the mathematical treatment of frequency-domain techniques. If we decompose a waveform into sinusoidal components, in much the same way that a prism separates white light into different colors, the sum of these sinusoidal components results in the original waveform. On the other hand, if any of these components is missing, the result is a different signal. In our treatment of frequency analysis, we will develop the proper mathematical tools (“prisms”) for the decomposition of signals (“light”) into sinusoidal frequency components (colors). Furthermore, the tools (“inverse prisms”) for synthesis of a given signal from its frequency components will also be developed. The basic motivation for developing the frequency analysis tools is to provide a mathematical and pictorial representation for the frequency components that are contained in any given signal. As in physics, the term spectrum is used when referring to the frequency content of a signal. The process of obtaining the spectrum of a given signal using the basic mathematical tools described in this chapter is known as frequency or spectral analysis. In contrast, the process of determining the spectrum of a signal in practice, based on actual measurements of the signal, is called spectrum estimation. This distinction is very important. In a practical problem the signal to be analyzed does not lend itself to an exact mathematical description. The signal is usually some information-bearing signal from which we are attempting to extract the relevant information. If the information that we wish to extract can be obtained either directly or indirectly from the spectral content of the signal, we can perform spectrum estimation on the information-bearing signal, and thus obtain an estimate of the signal spectrum. In fact, we can view spectral estimation as a type of spectral analysis performed on signals obtained from physical sources (e.g., speech, EEG, ECG, etc.). The instruments or software programs used to obtain spectral estimates of such signals are known as spectrum analyzers. Here, we will deal with spectral analysis. However, the subject of power spectrum estimation is beyond the scope of this chapter.

1.1 The Fourier Series for Continuous-Time Periodic Signals In this section we present the frequency analysis tools for continuous-time periodic signals. Examples of periodic signals encountered in practice are square waves, rectangular waves, triangular waves, and of course, sinusoids and complex exponentials. The basic mathematical representation of periodic signals is the Fourier series, which is a linear weighted sum of harmonically related sinusoids or complex exponentials. Jean Baptiste Joseph Fourier (1768–1830), a French mathematician, used such trigonometric series expansions in describing the phenomenon of heat conduction and temperature distribution through bodies. Although his work was motivated by the problem of heat conduction, the mathematical techniques that he developed during the early part of the nineteenth century now find application in a variety of problems encompassing many different fields, including optics, vibrations in mechanical systems, system theory, and electromagnetics.

232

Frequency Analysis of Signals

Recall that a linear combination of harmonically related complex exponentials of the form ∞ 

x(t) =

ck ej 2πkF0 t

(1.1)

k=−∞

is a periodic signal with fundamental period Tp = 1/F0 . Hence we can think of the exponential signals {ej 2πkF0 t , k = 0, ±1, ±2, . . .} as the basic “building blocks” from which we can construct periodic signals of various types by proper choice of the fundamental frequency and the coefficients {ck }. F0 determines the fundamental period of x(t) and the coefficients {ck } specify the shape of the waveform. Suppose that we are given a periodic signal x(t) with period Tp . We can represent the periodic signal by the series (1.1), called a Fourier series, where the fundamental frequency F0 is selected to be the reciprocal of the given period Tp . To determine the expression for the coefficients {ck }, we first multiply both sides of (1.1) by the complex exponential e−j 2πF0 lt where l is an integer and then integrate both sides of the resulting equation over a single period, say from 0 to Tp , or more generally, from t0 to t0 + Tp , where t0 is an arbitrary but mathematically convenient starting value. Thus we obtain 

t0 +Tp

x(t)e

−j 2π lF0 t

 dt =



t0 +Tp

e

t0

−j 2πlF0 t

t0

∞ 

 ck e

+j 2πkF0 t

dt

(1.2)

k=−∞

To evaluate the integral on the right-hand side of (1.2), we interchange the order of the summation and integration and combine the two exponentials. Hence ∞  k=−∞



t0 +Tp

ck

e

j 2π F0 (k−l)t

dt =

t0



∞ 

ck

k=−∞

ej 2πF0 (k−l)t j 2π F0 (k − l)

t0 +Tp (1.3) t0

For k = l , the right-hand side of (1.3) evaluated at the lower and upper limits, t 0 and t0 + Tp , respectively, yields zero. On the other hand, if k = l , we have 

t0 +Tp

t0

t0 +Tp   dt = t  = Tp  t0

Consequently, (1.2) reduces to 

t0 +Tp

x(t)e−j 2πlF0 t dt = cl Tp

t0

233

Frequency Analysis of Signals

and therefore the expression for the Fourier coefficients in terms of the given periodic signal becomes  t0 +Tp 1 cl = x(t)e−j 2πlF0 t dt Tp t0 Since t0 is arbitrary, this integral can be evaluated over any interval of length Tp , that is, over any interval equal to the period of the signal x(t). Consequently, the integral for the Fourier series coefficients will be written as  1 x(t)e−j 2πlF0 t dt cl = (1.4) Tp Tp An important issue that arises in the representation of the periodic signal x(t) by the Fourier series is whether or not the series converges to x(t) for every value of t , that is, whether the signal x(t) and its Fourier series representation ∞ 

ck ej 2πkF0 t

(1.5)

k=−∞

are equal at every value of t . The so-called Dirichlet conditions guarantee that the series (1.5) will be equal to x(t), except at the values of t for which x(t) is discontinuous. At these values of t , (1.5) converges to the midpoint (average value) of the discontinuity. The Dirichlet conditions are: 1. The signal x(t) has a finite number of discontinuities in any period. 2. The signal x(t) contains a finite number of maxima and minima during any period. 3. The signal x(t) is absolutely integrable in any period, that is,  |x(t)| dt < ∞

(1.6)

Tp

All periodic signals of practical interest satisfy these conditions. The weaker condition, that the signal has finite energy in one period,  |x(t)|2 dt < ∞

(1.7)

Tp

guarantees that the energy in the difference signal e(t) = x(t) −

∞ 

ck ej 2πkF0 t

k=−∞

is zero, although x(t) and its Fourier series may not be equal for all values of t . Note that (1.6) implies (1.7), but not vice versa. Also, both (1.7) and the Dirichlet

234

Frequency Analysis of Signals

conditions are sufficient but not necessary conditions (i.e., there are signals that have a Fourier series representation but do not satisfy these conditions). In summary, if x(t) is periodic and satisfies the Dirichlet conditions, it can be represented in a Fourier series as in (1.1), where the coefficients are specified by (1.4). These relations are summarized below. Frequency Analysis of Continuous-Time Periodic Signals Synthesis equation

∞ 

x(t) =

ck ej 2πkF0 t

(1.8)

x(t)e−j 2πkF0 t dt

(1.9)

k=−∞

Analysis equation

ck =

1 Tp

 Tp

In general, the Fourier coefficients ck are complex valued. Moreover, it is easily shown that if the periodic signal is real, ck and c−k are complex conjugates. As a result, if ck = |ck |ej θk then

c−k = |ck |−j θk

Consequently, the Fourier series may also be represented in the form x(t) = c0 + 2

∞ 

|ck | cos(2π kF0 t + θk )

(1.10)

k=1

where c0 is real valued when x(t) is real. Finally, we should indicate that yet another form for the Fourier series can be obtained by expanding the cosine function in (1.10) as cos(2πkF0 t + θk ) = cos 2π kF0 t cos θk − sin 2π kF0 t sin θk Consequently, we can rewrite (1.10) in the form x(t) = a0 +

∞ 

(ak cos 2π kF0 t − bk sin 2π kF0 t)

(1.11)

k=1

where a0 = c0 ak = 2|ck | cos θk bk = 2|ck | sin θk The expressions in (1.8), (1.10), and (1.11) constitute three equivalent forms for the Fourier series representation of a real periodic signal.

235

Frequency Analysis of Signals

1.2

Power Density Spectrum of Periodic Signals

A periodic signal has infinite energy and a finite average power, which is given as  1 Px = |x(t)|2 dt (1.12) Tp Tp If we take the complex conjugate of (1.8) and substitute for x∗ (t) in (1.12), we obtain  ∞  1 x(t) ck∗ e−j 2πkF0 t dt Px = Tp Tp k=−∞

=

∞ 

 ck∗

k=−∞

=

∞ 

1 Tp





x(t)e−j 2πkF0 t dt

(1.13)

Tp

|ck |2

k=−∞

Therefore, we have established the relation Px =

1 Tp

 |x(t)|2 dt = Tp

∞ 

|ck |2

(1.14)

k=−∞

which is called Parseval’s relationfor power signals. To illustrate the physical meaning of (1.14), suppose that x(t) consists of a single complex exponential x(t) = ck ej 2πkF0 t In this case, all the Fourier series coefficients except ck are zero. Consequently, the average power in the signal is Px = |ck |2 It is obvious that |ck |2 represents the power in the kth harmonic component of the signal. Hence the total average power in the periodic signal is simply the sum of the average powers in all the harmonics. If we plot the |ck |2 as a function of the frequencies kF0 , k = 0, ±1, ±2, . . . , the diagram that we obtain shows how the power of the periodic signal is distributed among the various frequency components. This diagram, which is illustrated in Fig. 1.2, is called the power density spectrum 1 of the periodic signal x(t). Since the power in a periodic signal exists only at discrete values of frequencies (i.e., F = 0, ±F0 , ±2F0 , . . .), the signal is said to have a line spectrum. The spacing between two consecutive spectral lines is equal to the reciprocal of the fundamental period Tp , whereas the shape of the spectrum (i.e., the power distribution of the signal), depends on the time-domain characteristics of the signal. 1

236

This function is also called the power spectral density or, simply, the power spectrum.

Frequency Analysis of Signals

|ck|2

Power density spectrum



… −4F0 −3F0 −2F0 −F0

0

F0 2F0 3F0 4F0

Frequency, F

Figure 1.2 Power density spectrum of a continuous-time periodic

signal.

As indicated in the preceding section, the Fourier series coefficients {ck } are complex valued, that is, they can be represented as ck = |ck |ej θk where θk = ⭿ck Instead of plotting the power density spectrum, we can plot the magnitude voltage spectrum {|ck |} and the phase spectrum {θk } as a function of frequency. Clearly, the power spectral density in the periodic signal is simply the square of the magnitude spectrum. The phase information is totally destroyed (or does not appear) in the power spectral density. If the periodic signal is real valued, the Fourier series coefficients {ck } satisfy the condition c−k = ck∗ Consequently, |ck |2 = |ck∗ |2 . Hence the power spectrum is a symmetric function of frequency. This condition also means that the magnitude spectrum is symmetric (even function) about the origin and the phase spectrum is an odd function. As a consequence of the symmetry, it is sufficient to specify the spectrum of a real periodic signal for positive frequencies only. Furthermore, the total average power can be expressed as Px = c02 + 2

∞ 

|ck |2

(1.15)

1 2 (ak + bk2 ) 2

(1.16)

k=1 ∞

= a02 +

k=1

which follows directly from the relationships given in Section 1.1 among {a k}, {b k}, and {ck } coefficients in the Fourier series expressions.

237

Frequency Analysis of Signals

x(t) A …



Figure 1.3

−Tp

Continuous-time periodic train of rectangular pulses.



τ 2

0

τ 2

Tp

t

EXAMPLE 1.1 Determine the Fourier series and the power density spectrum of the rectangular pulse train signal illustrated in Fig 1.3. Solution. The signal is periodic with fundamental period Tp and, clearly, satisfies the Dirichlet conditions. Consequently, we can represent the signal in the Fourier series given by (1.8) with the Fourier coefficients specified by (1.9). Since x(t) is an even signal [i.e., x(t) = x(−t)], it is convenient to select the integration interval from −Tp /2 to Tp /2. Thus (1.9) evaluated for k = 0 yields c0 =

1 Tp



Tp /2

−Tp /2

x(t) dt =

1 Tp



τ/2 −τ/2

A dt =

Aτ Tp

(1.17)

The term c0 represents the average value (dc component) of the signal x(t). For k = 0 we have  τ/2  τ/2 1 A e−j 2π F0 kt Ae−j 2π kF0 t dt = ck = Tp −τ/2 Tp −j 2πkF0 −τ/2 =

A ej π kF0 τ − e−j π kF0 τ πF0 kTp j2

=

Aτ sin πkF0 τ , Tp πkF0 τ

(1.18)

k = ±1, ±2, . . .

It is interesting to note that the right-hand side of (1.18) has the form (sin φ)/φ , where φ = πkF0 τ . In this case φ takes on discrete values since F0 and τ are fixed and the index k varies. However, if we plot (sin φ)/φ with φ as a continuous parameter over the range −∞ < φ < ∞, we obtain the graph shown in Fig 1.4. We observe that this function decays to zero as φ → ±∞, has a maximum value of unity at φ = 0, and is zero at multiples of π (i.e., at φ = mπ , m = ±1, ±2, . . .). It is clear that the Fourier coefficients given by (1.18) are the sample values of the (sin φ)/φ function for φ = πkF0 τ and scaled in amplitude by Aτ/Tp .

1

0

−7π −6π −5π −4π −3π −2π −π 0

Figure 1.4 The function (sin φ)/φ .

238

sin φ φ

π













φ

Frequency Analysis of Signals

Since the periodic function x(t) is even, the Fourier coefficients ck are real. Consequently, the phase spectrum is either zero, when ck is positive, or π when ck is negative. Instead of plotting the magnitude and phase spectra separately, we may simply plot {ck } on a single graph, showing both the positive and negative values ck on the graph. This is commonly done in practice when the Fourier coefficients {ck } are real. Figure 1.5 illustrates the Fourier coefficients of the rectangular pulse train when Tp is fixed and the pulse width τ is allowed to vary. In this case Tp = 0.25 second, so that F0 = 1/Tp = 4 Hz and τ = 0.05Tp , τ = 0.1Tp , and τ = 0.2Tp . We observe that the effect of decreasing τ while keeping Tp fixed is to spread out the signal power over the frequency range. The spacing between adjacent spectral lines is F0 = 4 Hz, independent of the value of the pulse width τ . On the other hand, it is also instructive to fix τ and vary the period Tp when Tp > τ . Figure 1.6 illustrates this condition when Tp = 5τ , Tp = 10τ , and Tp = 20τ . In this case, the spacing between adjacent spectral lines decreases as Tp increases. In the limit as Tp → ∞, the Fourier coefficients ck approach zero due to the factor of Tp in the denominator of (1.18). This behavior is consistent with the fact that as Tp → ∞ and τ remains fixed, the resulting signal is no longer a power signal. Instead, it becomes an energy signal and its average power is zero. The spectra of finite energy signals are described in the next section. We also note that if k = 0 and sin(πkF0 τ ) = 0, then ck = 0. The harmonics with zero power occur at frequencies kF0 such that π(kF0 )τ = mπ , m = ±1, ±2, . . . , or at kF0 = m/τ . For example, if F0 = 4 Hz and τ = 0.2Tp , it follows that the spectral components at ±20 Hz, ±40 Hz, . . . have zero power. These frequencies correspond to the Fourier coefficients ck , k = ±5, ±10, ±15, . . . . On the other hand, if τ = 0.1Tp , the spectral components with zero power are k = ±10, ±20, ±30, . . . . The power density spectrum for the rectangular pulse train is    Aτ 2   , k=0  T |ck |2 =  p 2  (1.19) 2  Aτ sin πkF0 τ   , k = ±1, ±2, . . .  Tp πkF0 τ ck τ = 0.2Tp

F

ck

τ = 0.1Tp

F

Figure 1.5

Fourier coefficients of the rectangular pulse train when Tp is fixed and the pulse width τ varies.

ck

0

τ = 0.05Tp

F

239

Frequency Analysis of Signals

ck

Tp = 5τ

F

0 ck

Tp = 10τ

F

0

Figure 1.6

Fourier coefficient of a rectangular pulse train with fixed pulse width τ and varying period Tp .

ck

Tp = 20τ

F

0

1.3 The Fourier Transform for Continuous-Time Aperiodic Signals In Section 1.1 we developed the Fourier series to represent a periodic signal as a linear combination of harmonically related complex exponentials. As a consequence of the periodicity, we saw that these signals possess line spectra with equidistant lines. The line spacing is equal to the fundamental frequency, which in turn is the inverse of the fundamental period of the signal. We can view the fundamental period as providing the number of lines per unit of frequency (line density), as illustrated in Fig 1.6. With this interpretation in mind, it is apparent that if we allow the period to increase without limit, the line spacing tends toward zero. In the limit, when the period becomes infinite, the signal becomes aperiodic and its spectrum becomes continuous. This argument suggests that the spectrum of an aperiodic signal will be the envelope of the line spectrum in the corresponding periodic signal obtained by repeating the aperiodic signal with some period Tp . Let us consider an aperiodic signal x(t) with finite duration as shown in Fig 1.7(a). From this aperiodic signal, we can create a periodic signal xp (t) with period Tp , as shown in Fig 1.7(b). Clearly, x (t) p = x(t) in the limit as T p → ∞, that is, x(t) = lim xp (t) Tp →∞

This interpretation implies that we should be able to obtain the spectrum of x(t) from the spectrum of xp (t) simply by taking the limit as Tp → ∞. We begin with the Fourier series representation of xp (t), xp (t) =

∞  k=−∞

240

ck ej 2πkF0 t ,

F0 =

1 Tp

(1.20)

Frequency Analysis of Signals

x(t)

−Tp /2

0

Tp /2

t

(a) x(t) …

Figure 1.7



(a) Aperiodic signal x(t) and (b) periodic signal xp (t) constructed by repeating x(t) with a period Tp .

−Tp

−Tp /2

0

Tp /2

Tp/2

t

(b)

where ck =

1 Tp



Tp /2

xp (t)e−j 2πkF0 t dt

−Tp /2

(1.21)

Since xp (t) = x(t) for −Tp /2 ≤ t ≤ Tp /2, (1.21) can be expressed as ck =

1 Tp



Tp /2

x(t)e−j 2πkF0 t dt

−Tp /2

(1.22)

It is also true that x(t) = 0 for |t| > Tp /2. Consequently, the limits on the integral in (1.22) can be replaced by −∞ and ∞. Hence ck =

1 Tp





x(t)e−j 2πkF0 t dt

−∞

(1.23)

Let us now define a function X(F ), called the Fourier transform of x(t), as  X(F ) =



x(t)e−j 2πF t dt

−∞

(1.24)

X(F ) is a function of the continuous variable F . It does not depend on Tp or F0 . However, if we compare (1.23) and (1.24), it is clear that the Fourier coefficients ck can be expressed in terms of X(F ) as ck =

1 X(kF0 ) Tp

or equivalently,

 Tp ck = X(kF0 ) = X

k Tp

 (1.25)

241

Frequency Analysis of Signals

Thus the Fourier coefficients are samples of X(F ) taken at multiples of F0 and scaled by F0 (multiplied by 1/Tp ). Substitution for ck from (1.25) into (1.20) yields   ∞ 1  k xp (t) = ej 2πkF0 t X Tp Tp

(1.26)

k=−∞

We wish to take the limit of (1.26) as Tp approaches infinity. First, we define F = 1/Tp . With this substitution, (1.26) becomes xp (t) =

∞ 

X(kF )ej 2πk F t F

(1.27)

k=−∞

It is clear that in the limit as Tp approaches infinity, xp (t) reduces to x(t). Also, F becomes the differential dF and k F becomes the continuous frequency variable F . In turn, the summation in (1.27) becomes an integral over the frequency variable F . Thus lim xp (t) = x(t) = lim

Tp →∞

F →0

 x(t) =



∞ 

X(kF )e−j 2πk F t F

k=−∞

(1.28)

X(F )ej 2πF t dF

−∞

This integral relationship yields x(t) when X(F ) is known, and it is called the inverse Fourier transform. This concludes our heuristic derivation of the Fourier transform pair given by (1.24) and (1.28) for an aperiodic signal x(t). Although the derivation is not mathematically rigorous, it led to the desired Fourier transform relationships with relatively simple intuitive arguments. In summary, the frequency analysis of continuoustime aperiodic signals involves the following Fourier transform pair. Frequency Analysis of Continuous-Time Aperiodic Signals  Synthesis equation (inverse transform)

x(t) =

Analysis equation (direct transform)

X(F ) =



X(F )ej 2πF t dF

−∞





−∞

x(t)e−j 2πF t dt

(1.29)

(1.30)

It is apparent that the essential difference between the Fourier series and the Fourier transform is that the spectrum in the latter case is continuous and hence the synthesis of an aperiodic signal from its spectrum is accomplished by means of integration instead of summation.

242

Frequency Analysis of Signals

Finally, we wish to i ndicate that the Fourier transform pair in (1.29) and (1.30) can be expressed in terms of the radian frequency variable  = 2π F . Since dF = d/2π , (1.29) and (1.30) become  ∞ 1 x(t) = X()ej t d 2π −∞  ∞ X() = x(t)e−j t dt −∞

(1.31) (1.32)

The set of conditions that guarantee the existence of the Fourier transform is the Dirichlet conditions, which may be expressed as: 1. The signal x(t) has a finite number of finite discontinuities. 2. The signal x(t) has a finite number of maxima and minima. 3. The signal x(t) is absolutely integrable, that is, 

∞ −∞

|x(t)| dt < ∞

(1.33)

The third condition follows easily from the definition of the Fourier transform, given in (1.30). Indeed,   |X(F )| = 



x(t)e

−j 2πF t

−∞

   dt  ≤



−∞

|x(t)| dt

Hence |X(F )| < ∞ if (1.33) is satisfied. A weaker condition for the existence of the Fourier transform is that x(t) has finite energy; that is,  ∞ (1.34) |x(t)|2 dt < ∞ −∞

Note that if a signal x(t) is absolutely integrable, it will also have finite energy. That is, if  ∞ |x(t)| dt < ∞ −∞

then

 Ex =



−∞

|x(t)|2 dt < ∞

(1.35)

However, the converse is not true. That is, a signal may have finite energy but may not be absolutely integrable. For example, the signal x(t) =

sin 2π F0 t πt

(1.36)

243

Frequency Analysis of Signals

is square integrable but is not absolutely integrable. This signal has the Fourier transform  1, |F | ≤ F0 X(F ) = (1.37) 0, |F | > F0 Since this signal violates (1.33), it is apparent that the Dirichlet conditions are sufficient but not necessary for the existence of the Fourier transform. In any case, nearly all finite energy signals have a Fourier transform, so that we need not worry about the pathological signals, which are seldom encountered in practice.

1.4

Energy Density Spectrum of Aperiodic Signals

Let x(t) be any finite energy signal with Fourier transform X(F ). Its energy is  ∞ |x(t)|2 dt Ex = −∞

which, in turn, may be expressed in terms of X(F ) as follows:  ∞ Ex = x(t)x ∗ (t) dt −∞

 =







X ∗ (F )e−j 2πF t dF

x(t) dt 

=  =

−∞ ∞

−∞

X ∗ (F ) dF



−∞ ∞ −∞



x(t)e−j 2πF t dt

 

−∞

|X(F )|2 dF

Therefore, we conclude that  Ex =



−∞

 |x(t)|2 dt =



−∞

|X(F )|2 dF

(1.38)

This is Parseval’s relation for aperiodic, finite energy signals and expresses the principle of conservation of energy in the time and frequency domains. The spectrum X(F ) of a signal is, in general, complex valued. Consequently, it is usually expressed in polar form as X(F ) = |X(F )|ej (F ) where |X(F )| is the magnitude spectrum and (F ) is the phase spectrum, (F ) = ⭿X(F ) On the other hand, the quantity Sxx (F ) = |X(F )|2

244

(1.39)

Frequency Analysis of Signals

which is the integrand in (1.38), represents the distribution of energy in the signal as a function of frequency. Hence Sxx (F ) is called the energy density spectrum of x(t). The integral of Sxx (F ) over all frequencies gives the total energy in the signal. Viewed in another way, the energy in the signal x(t) over a band of frequencies F1 ≤ F ≤ F1 + F is  F1 +F Sxx (F ) dF ≥ 0 F1

which implies that Sxx (f ) ≥ 0 for all F . From (1.39) we observe that Sxx (F ) does not contain any phase information [i.e., Sxx (F ) is purely real and nonnegative]. Since the phase spectrum of x(t) is not contained in Sxx (F ), it is impossible to reconstruct the signal given Sxx (F ). Finally, as in the case of Fourier series, it is easily shown that if the signal x(t) is real, then |X(−F )| = |X(F )|

(1.40)

⭿X(−F ) = −⭿X(F )

(1.41)

By combining (1.40) and (1.39), we obtain Sxx (−F ) = Sxx (F )

(1.42)

In other words, the energy density spectrum of a real signal has even symmetry. EXAMPLE 1.2 Determine the Fourier transform and the energy density spectrum of a rectangular pulse signal defined as  A, |t| ≤ τ/2 x(t) = (1.43) 0, |t| > τ/2 and illustrated in Fig 1.8(a). Solution. Clearly, this signal is aperiodic and satisfies the Dirichlet conditions. Hence its Fourier transform exists. By applying (1.30), we find that  X(F ) =

τ/2

−τ/2

Ae−j 2π F t dt = Aτ

sin πF τ πF τ

(1.44)

We observe that X(F ) is real and hence it can be depicted graphically using only one diagram, as shown in Fig 1.8(b). Obviously, X(F ) has the shape of the (sin φ)/φ function shown in Fig 1.4. Hence the spectrum of the rectangular pulse is the envelope of the line spectrum (Fourier coefficients) of the periodic signal obtained by periodically repeating the pulse with period Tp as in Fig 1.3. In other words, the Fourier coefficients ck in the corresponding periodic signal xp (t) are simply samples of X(F ) at frequencies kF0 = k/Tp . Specifically,   1 1 k X(kF0 ) = X ck = (1.45) Tp Tp Tp From (1.44) we note that the zero crossings of X(F ) occur at multiples of 1/τ . Furthermore, the width of the main lobe, which contains most of the signal energy, is equal to 2/τ . As the

245

Frequency Analysis of Signals

x(t) A



τ 2

τ 2

0

t

(a) X(F) Aτ

Figure 1.8

(a) Rectangular pulse and (b) its Fourier transform.

−2 τ

−1 τ

0

2 τ

1 τ

F

(b)

pulse duration τ decreases (increases), the main lobe becomes broader (narrower) and more energy is moved to the higher (lower) frequencies, as illustrated in Fig 1.9. Thus as the signal X(F)

x(t) A Aτ −

τ 0 τ 2 2

t



0

1 τ

x(t)

τ 2

F

X(F)

A



1 τ



0

τ 2

t



1 τ

0

1 τ

F

Aτ X(F) x(t)

Figure 1.9

Fourier transform of a rectangular pulse for various width values.

246

A



τ 2

0

τ 2

t

0 1 1 − τ τ

F

Frequency Analysis of Signals

pulse is expanded (compressed) in time, its transform is compressed (expanded) in frequency. This behavior, between the time function and its spectrum, is a type of uncertainty principle that appears in different forms in various branches of science and engineering. Finally, the energy density spectrum of the rectangular pulse is  Sxx (F ) = (Aτ )2

2

sin πF τ πF τ

2 (1.46)

Frequency Analysis of Discrete-Time Signals In Section 1 we developed the Fourier series representation for continuous-time periodic (power) signals and the Fourier transform for finite energy aperiodic signals. In this section we repeat the development for the class of discrete-time signals. As we have observed from the discussion of Section 1, the Fourier series representation of a continuous-time periodic signal can consist of an infinite number of frequency components, where the frequency spacing between two successive harmonically related frequencies is 1/Tp , and where Tp is the fundamental period. Since the frequency range for continuous-time signals extends from −∞ to ∞, it is possible to have signals that contain an infinite number of frequency components. In contrast, the frequency range for discrete-time signals is unique over the interval (−π, π ) or (0, 2π). A discrete-time signal of fundamental period N can consist of frequency components separated by 2π/N radians or f = 1/N cycles. Consequently, the Fourier series representation of the discrete-time periodic signal will contain at most N frequency components. This is the basic difference between the Fourier series representations for continuous-time and discrete-time periodic signals.

2.1

The Fourier Series for Discrete-Time Periodic Signals

Suppose that we are given a periodic sequence x(n) with period N , that is, x(n) = x(n + N ) for all n. The Fourier series representation for x(n) consists of N harmonically related exponential functions ej 2πkn/N ,

k = 0, 1, . . . , N − 1

and is expressed as x(n) =

N−1 

ck ej 2πkn/N

(2.1)

k=0

where the {ck } are the coefficients in the series representation. To derive the expression for the Fourier coefficients, we use the following formula:  N−1  N, k = 0, ±N, ±2N, . . . j 2πkn/N (2.2) e = 0, otherwise n=0

247

Frequency Analysis of Signals

Note the similarity of (2.2) with the continuous-time counterpart in (1.3). The proof of (2.2) follows immediately from the application of the geometric summation formula  N−1 N, a=1  n (2.3) a = 1 − aN , a = 1 n=0 1−a The expression for the Fourier coefficients ck can be obtained by multiplying both sides of (2.1) by the exponential e −j 2πln/N and summing the product from n = 0 to n = N − 1. Thus N−1 

x(n)e−j 2πln/N =

n=0

N−1  N−1 

ck ej 2π(k−l)n/N

(2.4)

n=0 k=0

If we perform the summation over n first, in the right-hand side of (2.4), we obtain  N−1  N, k − l = 0, ±N, ±2N, . . . j 2π(k−l)n/N (2.5) e = 0, otherwise n=0

where we have made use of (2.2). Therefore , the right-hand side of (2.4) reduces to N cl and hence cl =

N−1 1  x(n)e−j 2πln/N , N

l = 0, 1, . . . , N − 1

(2.6)

n=0

Thus we have the desired expression for the Fourier coefficients in terms of the signal x(n). The relationships (2.1) and (2.6) for the frequency analysis of discrete-time signals are summarized below. Frequency Analysis of Discrete-Time Periodic Signals

Synthesis equation

x(n) =

N−1 

ck ej 2πkn/N

(2.7)

k=0

Analysis equation

ck =

N−1 1  x(n)e−j 2πkn/N N

(2.8)

n=0

Equation (2.7) is often called the discrete-time Fourier series (DTFS). The Fourier coefficients {ck }, k = 0, 1, . . . , N − 1 provide the description of x(n) in the frequency domain, in the sense that ck represents the amplitude and phase associated with the frequency component sk (n) = ej 2πkn/N = ej ωk n where ωk = 2πk/N .

248

Frequency Analysis of Signals

Recall that the functions sk(n) are periodic with period N. Hence sk(n) = sk(n + N). In view of this periodicity, it follows that the Fourier coefficients ck , when viewed beyond the range k = 0, 1, . . . , N − 1, also satisfy a periodicity condition. Indeed, from (2.8), which holds for every value of k, we have ck+N =

N−1 N−1 1  1  x(n)e−j 2π(k+N)n/N = x(n)e−j 2πkn/N = ck N N n=0

(2.9)

n=0

Therefore, the Fourier series coefficients {ck } form a periodic sequence when extended outside of the range k = 0, 1, . . . , N − 1. Hence ck+N = ck that is, {ck } is a periodic sequence with fundamental period N . Thus the spectrum of a signal x(n) , which is periodic with period N , is a periodic sequence with period N . Consequently, any N consecutive samples of the signal or its spectrum provide a complete description of the signal in the time or frequency domains. Although the Fourier coefficients form a periodic sequence, we will focus our attention on the single period with range k = 0, 1, . . . , N − 1. This is convenient, since in the frequency domain, this amounts to covering the fundamental range 0 ≤ ωk = 2π k/N < 2π , for 0 ≤ k ≤ N − 1. In contrast, the frequency range −π < ωk = 2πk/N ≤ π corresponds to −N/2 < k ≤ N/2, which creates an inconvenience when N is odd. Clearly, if we use a sampling frequency Fs , the range 0 ≤ k ≤ N − 1 corresponds to the frequency range 0 ≤ F < Fs . EXAMPLE 2.1 Determine the spectra of the signals √ (a) x(n) = cos 2πn (b) x(n) = cos πn/3 (c) x(n) is periodic with period N = 4 and x(n) = {1, 1, 0, 0} ↑

Solution.

√ √ (a) For ω0 = 2π , we have f0 = 1/ 2. Since f0 is not a rational number, the signal is not periodic. Consequently, this signal cannot be expanded in a Fourier series. Nevertheless, the signal does possess a√ spectrum. Its spectral content consists of the single frequency component at ω = ω0 = 2π .

(b) In this case f0 = (2.8) we have

1 6

and hence x(n) is periodic with fundamental period N = 6. From

1 x(n)e−j 2π kn/6 , 6 5

ck =

k = 0, 1, . . . , 5

n=0

249

Frequency Analysis of Signals

However, x(n) can be expressed as

x(n) = cos

1 1 2πn = ej 2π n/6 + e−j 2π n/6 6 2 2

which is already in the form of the exponential Fourier series in (2.7). In comparing the two exponential terms in x(n) with (2.7), it is apparent that c1 = 21 . The second exponential in x(n) corresponds to the term k = −1 in (2.7). However, this term can also be written as e−j 2π n/6 = ej 2π(5−6)n/6 = ej 2π(5n)/6 which means that c−1 = c5 . But this is consistent with (2.9), and with our previous observation that the Fourier series coefficients form a periodic sequence of period N . Consequently, we conclude that c0 = c2 = c3 = c4 = 0 c1 =

1 , 2

c5 =

1 2

(c) From (2.8), we have

ck =

3 1 x(n)e−j 2π kn/4 , 4

k = 0, 1, 2, 3

n=0

or ck =

1 (1 + e−j π k/2 ), 4

k = 0, 1, 2, 3

For k = 0, 1, 2, 3 we obtain c0 =

1 , 2

c1 =

1 (1 − j ), 4

c2 = 0,

c3 =

1 (1 + j ) 4

The magnitude and phase spectra are

|c0 | =

1 , 2

⭿c0 = 0,

√ 2 , 4 π ⭿c1 = − , 4 |c1 | =

|c2 | = 0,

|c3 | =

⭿c2 = undefined,

√ 2 4 ⭿c3 =

Figure 2.1 illustrates the spectral content of the signals in (b) and (c).

250

π 4

Frequency Analysis of Signals

ck 1 2 …

… −5

−1 0 1 2 3 4 5 6

k

(a) |ck| 1 2 …

… −3 −2 −1 0 1 2 3 4

k

(b) ck π 4

… −3



1

5

−2 −1 0 −

Figure 2.1

k

π 4

Spectra of the periodic signals discussed in Example 2.1 (b) and (c).

2.2

2 3 4

(c)

Power Density Spectrum of Periodic Signals

The average power of a discrete-time periodic signal with period N is defined as Px =

N−1 1  |x(n)|2 N

(2.10)

n=0

We shall now derive an expression for Px in terms of the Fourier coefficient {ck }. If we use the relation (2.7) in (2.10), we have Px =

N−1 1  x(n)x ∗ (n) N n=0

 N−1 N−1  1  ∗ −j 2πkn/N = x(n) ck e N n=0

k=0

251

Frequency Analysis of Signals

Now, we can interchange the order of the two summations and make use of (2.8), obtaining  N−1 N−1  1  ∗ −j 2πkn/N Px = ck x(n)e N k=0 n=0 (2.11) N−1 N−1   1 = |ck |2 = |x(n)|2 N k=0

n=0

which is the desired expression for the average power in the periodic signal. In other words, the average power in the signal is the sum of the powers of the individual frequency components. We view (2.11) as a Parseval’s relation for discrete-time periodic signals. The sequence |ck |2 for k = 0, 1, . . . , N − 1 is the distribution of power as a function of frequency and is called the power density spectrum of the periodic signal. If we are interested in the energy of the sequence x(n) over a single period, (2.11) implies that N−1 N−1   EN = |x(n)|2 = N |ck |2 (2.12) n=0

k=0

which is consistent with our previous results for continuous-time periodic signals. If the signal x(n) is real [i.e., x ∗ (n) = x(n)], then, proceeding as in Section 2.1, we can easily show that ck∗ = c−k (2.13) or equivalently, |c−k | = |ck | −⭿c−k = ⭿ck

(even symmetry)

(2.14)

(odd symmetry)

(2.15)

These symmetry properties for the magnitude and phase spectra of a periodic signal, in conjunction with the periodicity property, have very important implications on the frequency range of discrete-time signals. Indeed, by combining (2.9) with (2.14) and (2.15), we obtain |ck | = |cN−k |

(2.16)

⭿ck = −⭿cN−k

(2.17)

and More specifically, we have |c0 | = |cN |, |c1 | = |cN−1 |, |cN/2 | = |cN/2 |, |c(N −1)/2 | = |c(N +1)/2 |,

252

⭿c0 = −⭿cN = 0 ⭿c1 = −⭿cN−1 ⭿cN/2 = 0 ⭿c(N −1)/2 = −⭿c(N +1)/2

if N is even if N is odd

(2.18)

Frequency Analysis of Signals

Thus, for a real signal, the spectrum ck , k = 0, 1, . . . , N/2 for N even, or k = 0, 1, . . . , (N − 1)/2 for N odd, completely specifies the signal in the frequency domain. Clearly, this is consistent with the fact that the highest relative frequency that can be represented by a discrete-time signal is equal to π . Indeed, if 0 ≤ ωk = 2π k/N ≤ π , then 0 ≤ k ≤ N/2. By making use of these symmetry properties of the Fourier series coefficients of a real signal, the Fourier series in (2.7) can also be expressed in the alternative forms x(n) = c0 + 2

L  k=1

= a0 +

2π |ck | cos kn + θk N

L   k=1



 (2.19)

2π 2π ak cos kn − bk sin kn N N

 (2.20)

where a0 = c0 , ak = 2|ck | cos θk , bk = 2|ck | sin θk , and L = N/2 if N is even and L = (N − 1)/2 if N is odd. Finally, we note that as in the case of continuous-time signals, the power density spectrum |ck |2 does not contain any phase information. Furthermore, the spectrum is discrete and periodic with a fundamental period equal to that of the signal itself. EXAMPLE 2.2

Periodic “Square-Wave” Signal

Determine the Fourier series coefficients and the power density spectrum of the periodic signal shown in Fig 2.2. Solution.

By applying the analysis equation (2.8) to the signal shown in Fig 2.2, we obtain

ck =

N−1 L−1 1  1  −j 2π kn/N x(n)e−j 2π kn/N = Ae , N N n=0

k = 0, 1, . . . , N − 1

n=0

which is a geometric summation. Now we can use (2.3) to simplify the summation above. Thus we obtain  AL  L−1 A  −j 2π k/N n  N , (e ) = ck = −j 2π kL/N  N  A 1−e n=0 , N 1 − e−j 2π k/N

k=0 k = 1, 2, . . . , N − 1 x(n) A

Figure 2.2

Discrete-time periodic square-wave signal.

−N

0

L

N

n

253

Frequency Analysis of Signals

The last expression can be simplified further if we note that 1 − e−j 2π kL/N e−j π kL/N ej π kL/N − e−j π kL/N = −j π k/N −j 2π k/N 1−e e ej π k/N − e−j π k/N = e−j π k(L−1)/N

Therefore,

 AL   , N ck = A sin(πkL/N)   e−j π k(L−1)/N , N sin(πk/N)

sin(πkL/N) sin(πk/N)

k = 0, +N, ±2N, . . . (2.21) otherwise

The power density spectrum of this periodic signal is  2   AL ,   N |ck |2 =  2    sin πkL/N 2 A   ,  N sin πk/N

k = 0, +N, ±2N, . . . (2.22) otherwise

Figure 2.3 illustrates the plots of |ck |2 for L = 2, N = 10 and 40, and A = 1.

2.3

The Fourier Transform of Discrete-Time Aperiodic Signals

Just as in the case of continuous-time aperiodic energy signals, the frequency analysis of discrete-time aperiodic finite-energy signals involves a Fourier transform of the time-domain signal. Consequently, the development in this section parallels, to a large extent, that given in Section 1.3. The Fourier transform of a finite-energy discrete-time signal x(n) is defined as

X(ω) =

∞ 

x(n)e−j ωn

(2.23)

n=−∞

Physically, X(ω) represents the frequency content of the signal x(n). In other words, X(ω) is a decomposition of x(n) into its frequency components. We observe two basic differences between the Fourier transform of a discretetime finite-energy signal and the Fourier transform of a finite-energy analog signal. First, for continuous-time signals, the Fourier transform, and hence the spectrum of the signal, have a frequency range of (−∞, ∞). In contrast, the frequency range for a discrete-time signal is unique over the frequency interval of (−π, π ) or, equivalently, (0, 2π). This property is reflected in the Fourier transform of the signal. Indeed, X(ω)

254

Frequency Analysis of Signals

L = 2, N = 10 5 4 3 Nc

k

| |

2 1 0 1 2 1. 5

1

0. 5 0 0.5 Frequency (Cycles/Sampling interval)

1

1.5

1

1.5

L = 2, N = 40 5 4 3 N|ck

|

2 1 0 1 2 1.5

1

0.5 0 0.5 Frequency (Cycles/Sampling interval)

Figure 2.3 Plot of the power density spectrum given by (2.22).

is periodic with period 2π , that is, X(ω + 2π k) =

∞ 

x(n)e−j (ω+2πk)n

n=−∞

=

∞ 

x(n)e−j ωn e−j 2πkn

(2.24)

n=−∞

=

∞ 

x(n)e−j ωn = X(ω)

n=−∞

Hence X(ω) is periodic with period 2π . But this property is just a consequence of the fact that the frequency range for any discrete-time signal is limited to (−π, π ) or

255

Frequency Analysis of Signals

(0, 2π), and any frequency outside this interval is equivalent to a frequency within the interval. The second basic difference is also a consequence of the discrete-time nature of the signal. Since the signal is discrete in time, the Fourier transform of the signal involves a summation of terms instead of an integral, as in the case of continuous-time signals. Since X(ω) is a periodic function of the frequency variable ω, it has a Fourier series expansion, provided that the conditions for the existence of the Fourier series, described previously, are satisfied. In fact, from the definition of the Fourier transform X(ω) of the sequence x(n), given by (2.23), we observe that X(ω) has the form of a Fourier series. The Fourier coefficients in this series expansion are the values of the sequence x(n). To demonstrate this point, let us evaluate the sequence x(n) from X(ω). First, we multiply both sides (2.23) by e j ωm and integrate over the interval (−π, π ). Thus we have  π  π  ∞ X(ω)ej ωm dω = x(n)e−j ωn ej ωm dω (2.25) −π

−π

n=−∞

The integral on the right-hand side of (2.25) can be evaluated if we can interchange the order of summation and integration. This interchange can be made if the series XN (ω) =

N 

x(n)e−j ωn

n=−N

converges uniformly to X(ω) as N → ∞. Uniform convergence means that, for every ω, XN (ω) → X(ω), as N → ∞. The convergence of the Fourier transform is discussed in more detail in the following section. For the moment, let us assume that the series converges uniformly, so that we can interchange the order of summation and integration in (2.25). Then 



π

ej ω(m−n) dω =

−π

2π, 0,

m=n m = n

Consequently, ∞  n=−∞





π

x(n)

e

j ω(m−n)

−π

dω =

2π x(m), 0,

m=n m = n

(2.26)

By combining (2.25) and (2.26), we obtain the desired result that x(n) =

1 2π



π

X(ω)ej ωn dω −π

(2.27)

If we compare the integral in (2.27) with (1.9), we note that this is just the expression for the Fourier series coefficient for a function that is periodic with period

256

Frequency Analysis of Signals

2π . The only difference between (1.9) and ( 2.27) is the sign on the exponent in the integrand, which is a consequence of our definition of the Fourier transform as given by (2.23). Therefore, the Fourier transform of the sequence x(n), defined by (2.23), has the form of a Fourier series expansion. In summary, the Fourier transform pair for discrete-time signals is as follows. Frequency Analysis of Discrete-Time Aperiodic Signals 1 x(n) = 2π

Synthesis equation (inverse transform)

X(ω) =

Analysis equation (direct transform)

 X(ω)ej ωn dω

(2.28)



∞ 

x(n)e−j ωn

(2.29)

n=−∞

2.4

Convergence of the Fourier Transform

In the derivation of the inverse transform given by (2.28), we assumed that the series N  x(n)e−j ωn (2.30) XN (ω) = n=−N

converges uniformly to X(ω), given in the integral of (2.25), as N → ∞. By uniform convergence we mean that for each ω,   lim supω |X(ω) − XN (ω)| = 0 (2.31) N→∞

Uniform convergence is guaranteed if x(n) is absolutely summable. Indeed, if ∞ 

|x(n)| < ∞

(2.32)

n=−∞

then

  ∞ ∞      −j ωn  x(n)e |x(n)| < ∞ |X(ω)| =  ≤  n=−∞ n=−∞

Hence (2.32) is a sufficient condition for the existence of the discrete-time Fourier transform. We note that this is the discrete-time counterpart of the third Dirichlet condition for the Fourier transform of continuous-time signals. The first two conditions do not apply due to the discrete-time nature of {x(n)}. Some sequences are not absolutely summable, but they are square summable. That is, they have finite energy Ex =

∞ 

|x(n)|2 < ∞

(2.33)

n=−∞

257

Frequency Analysis of Signals

which is a weaker condition than (2.32). We would like to define the Fourier transform of finite-energy sequences, but we must relax the condition of uniform convergence. For such sequences we can impose a mean-square convergence condition:  π |X(ω) − XN (ω)|2 dω = 0 (2.34) lim N→∞ −π

Thus the energy in the error X(ω) − XN (ω) tends toward zero, but the error |X(ω) − XN (ω)| does not necessarily tend to zero. In this way we can include finite-energy signals in the class of signals for which the Fourier transform exists. Let us consider an example from the class of finite-energy signals. Suppose that  X(ω) =

1, 0,

|ω| ≤ ωc ωc < |ω| ≤ π

(2.35)

The reader should remember that X(ω) is periodic with period 2π . Hence (2.35) represents only one period of X(ω). The inverse transform of X(ω) results in the sequence 1 x(n) = 2π 1 = 2π



π

X(ω)ej ωn dω −π



ωc

−ωc

ej ωn dω =

For n = 0, we have x(0) = Hence

1 2π



ωc

sin ωc n , πn

−ωc dω =

ω c   , π x(n) = ωc sin ωc n   , π ωc n

n = 0

ωc π

n=0 n = 0

(2.36)

This transform pair is illustrated in Fig 2.4. Sometimes, the sequence {x(n)} in (2.36) is expressed as x(n) =

sin ωc n , πn

−∞ < n < ∞

(2.37)

with the understanding that at n = 0, x(n) = ωc /π . We should emphasize, however, that (sin ωc n)/π n is not a continuous function, and hence L’Hospital’s rule cannot be used to determine x(0).

258

Frequency Analysis of Signals

x(n) ωc π



π ωc

π ωc

0

n

(a) X(ω)

1

−π

−ωc

0

ωc

π

(b)

Figure 2.4 Fourier transform pair in (2.35) and (2.36).

Now let us consider the determination of the Fourier transform of the sequence given by (2.37). The sequence {x(n)} is not absolutely summable. Hence the infinite series ∞ ∞   sin ωc n −j ωn −j ωn x(n)e = e (2.38) πn n=−∞ n=−∞ does not converge uniformly for all ω. However, the sequence {x(n)} has a finite energy Ex = ωc /π as will be shown in Section 4.3. Hence the sum in (2.38) is guaranteed to converge to the X(ω) given by (2.35) in the mean-square sense. To elaborate on this point, let us consider the finite sum

XN (ω) =

N  sin ωc n −j ωn e πn

(2.39)

n=−N

Figure 2.5 shows the function XN (ω) for several values of N . We note that there is a significant oscillatory overshoot at ω = ωc , independent of the value of N . As N increases, the oscillations become more rapid, but the size of the ripple remains the same. One can show that as N → ∞, the oscillations converge to the point of the discontinuity at ω = ωc , but their amplitude does not go to zero. However, (2.34) is satisfied, and therefore XN (ω) converges to X(ω) in the mean-square sense.

259

Frequency Analysis of Signals

Figure 2.5 Illustration of convergence of the Fourier transform

and the Gibbs phenomenon at the point of discontinuity.

The oscillatory behavior of the approximation XN (ω) to the function X(ω) at a point of discontinuity of X(ω) is called the Gibbs phenomenon. A similar effect is observed in the truncation of the Fourier series of a continuous-time periodic signal, given by the synthesis equation (1.8). For example, the truncation of the Fourier series for the periodic square-wave signal in Example 1.1 gives rise to the same oscillatory behavior in the finite-sum approximation of x(t). The Gibbs phenomenon is encountered again in the design of practical, discrete-time FIR systems.

2.5

Energy Density Spectrum of Aperiodic Signals

Recall that the energy of a discrete-time signal x(n) is defined as Ex =

∞  n=−∞

260

|x(n)|2

(2.40)

Frequency Analysis of Signals

Let us now express the energy Ex in terms of the spectral characteristic X(ω). First we have    π ∞ ∞   1 ∗ ∗ −j ωn Ex = x (n)x(n) = x(n) X (ω)e dω 2π −π n=−∞ n=−∞ If we interchange the order of integration and summation in the equation above, we obtain  ∞  π  1 ∗ −j ωn X (ω) x(n)e dω Ex = 2π −π n=−∞ =

1 2π



π

−π

|X(ω)|2 dω

Therefore, the energy relation between x(n) and X(ω) is Ex =

∞ 

|x(n)|2 =

n=−∞

1 2π



π

−π

|X(ω)|2 dω

(2.41)

This is Parseval’s relation for discrete-time aperiodic signals with finite energy. The spectrum X(ω) is, in general, a complex-valued function of frequency. It may be expressed as X(ω) = |X(ω)|ej (ω) (2.42) where (ω) = ⭿X(ω) is the phase spectrum and |X(ω)| is the magnitude spectrum. As in the case of continuous-time signals, the quantity Sxx (ω) = |X(ω)|2

(2.43)

represents the distribution of energy as a function of frequency, and it is called the energy density spectrum of x(n). Clearly, Sxx (ω) does not contain any phase information. Suppose now that the signal x(n) is real. Then it easily follows that X∗ (ω) = X(−ω)

(2.44)

or equivalently, |X(−ω)| = |X(ω)|,

(even symmetry)

(2.45)

and ⭿X(−ω) = −⭿X(ω),

(odd symmetry)

(2.46)

From (2.43) it also follows that Sxx (−ω) = Sxx (ω),

(even symmetry)

(2.47)

261

Frequency Analysis of Signals

From these symmetry properties we conclude that the frequency range of real discrete-time signals can be limited further to the range 0 ≤ ω ≤ π (i.e., one-half of the period). Indeed, if we know X(ω) in the range 0 ≤ ω ≤ π , we can determine it for the range −π ≤ ω < 0 using the symmetry properties given above. As we have already observed, similar results hold for discrete-time periodic signals. Therefore, the frequency-domain description of a real discrete-time signal is completely specified by its spectrum in the frequency range 0 ≤ ω ≤ π . Usually, we work with the fundamental interval 0 ≤ ω ≤ π or 0 ≤ F ≤ Fs /2, expressed in hertz. We sketch more than half a period only when required by the specific application. EXAMPLE 2.3 Determine and sketch the energy density spectrum Sxx (ω) of the signal x(n) = a n u(n),

−1 < a < 1

Solution. Since |a| < 1, the sequence x(n) is absolutely summable, as can be verified by applying the geometric summation formula, ∞ 

|x(n)| =

n=−∞

∞ 

|a|n =

n=0

1 1. Alternatively, if X(z) converges for |z| = 1, then ∞  x(n)e−j ωn (2.57) X(z)|z=ej ω ≡ X(ω) = n=−∞

Therefore, the Fourier transform can be viewed as the z-transform of the sequence evaluated on the unit circle. If X(z) does not converge in the region |z| = 1 [i.e., if the unit circle is not contained in the region of convergence of X(z)], the Fourier transform X(ω) does not exist. Figure 2.9 illustrates the relationship between X(z) and X(ω) for the rectangular sequence in Example 2.4, where A = 1 and L = 10. We should note that the existence of the z-transform requires that the sequence {x(n)r −n } be absolutely summable for some value of r , that is, ∞ 

|x(n)r −n | < ∞

(2.58)

n=−∞

Hence if (2.58) converges only for values of r > r 0 > 1, the z-transform exists, but the Fourier transform does not exist. This is the case, for example, for causal sequences of the form x(n) = a n u(n), where |a| > 1.

20 ln|X(z)|

15 10 5 0 2 1 0 1 Re(z) ω = ±π

π/2 ω =−

X(z)

2 2

0 Im(z)

2

X(ejω) Imz ω = π/2

0

|X(ω)|

10 8 6

ejω Unit Circle

1

1

ω Rez ω =0

4 2 0 −π

−π/2

0

π/2

Figure 2.9 relationship between X(z) and X(ω) for the sequence in Example 2.4, with A = 1 and L = 10

266

π

Frequency Analysis of Signals

There are sequences, however, that do not satisfy the requirement in (2.58), for example, the sequence x(n) =

sin ωc n , πn

−∞ < n < ∞

(2.59)

This sequence does not have a z-transform. Since it has a finite energy, its Fourier transform converges in the mean-square sense to the discontinuous function X(ω), defined as  1, |ω| < ωc X(ω) = (2.60) 0, ωc < |ω| ≤ π In conclusion, the existence of the z-transform requires that (2.58) be satisfied for some region in the z-plane. If this region contains the unit circle, the Fourier transform X(ω) exists. However, the existence of the Fourier transform, which is defined for finite energy signals, does not necessarily ensure the existence of the z-transform.

2.7

The Cepstrum

Let us consider a sequence {x(n)} having a z-transform X(z). We assume that {x(n)} is a stable sequence so that X(z) converges on the unit circle. The complex cepstrum of the sequence {x(n)} is defined as the sequence {cx (n)}, which is the inverse ztransform of Cx (z), where Cx (z) = ln X(z) (2.61) The complex cepstrum exists if Cx (z) converges in the annular region r1 < |z| < r2 , where 0 < r1 < 1 and r2 > 1. Within this region of convergence, Cx (z) can be represented by the Laurent series Cx (z) = ln X(z) =

∞ 

cx (n)z−n

(2.62)

n=−∞

where

1 cx (n) = 2πj

 ln X(z)zn−1 dz

(2.63)

C

C is a closed contour about the origin and lies within the region of convergence. Clearly, if Cx (z) can be represented as in (2.62), the complex cepstrum sequence {cx (n)} is stable. Furthermore, if the complex cepstrum exists, Cx (z) converges on the unit circle and hence we have Cx (ω) = ln X(ω) =

∞ 

cx (n)e−j ωn

(2.64)

n=−∞

where {cx (n)} is the sequence obtained from the inverse Fourier transform of ln X(ω), that is,  π 1 cx (n) = ln X(ω)ej ωn dω (2.65) 2π −π

267

Frequency Analysis of Signals

If we express X(ω) in terms of its magnitude and phase, say X(ω) = |X(ω)|ej θ(ω)

(2.66)

ln X(ω) = ln |X(ω)| + j θ (ω)

(2.67)

then By substituting (2.67) into (2.65), we obtain the complex cepstrum in the form  π 1 cx (n) = [ln |X(ω)| + j θ (ω)]ej ωn dω (2.68) 2π −π We can separate the inverse Fourier transform in (2.68) into the inverse Fourier transforms of ln |X(ω)| and θ(ω):  π 1 cm (n) = ln |X(ω)|ej ωn dω (2.69) 2π −π  π 1 θ(ω)ej ωn dω cθ (n) = (2.70) 2π −π In some applications, such as speech signal processing, only the component cm (n) is computed. In such a case the phase of X(ω) is ignored. Therefore, the sequence {x(n)} cannot be recovered from {cm (n)}. That is, the transformation from {x(n)} to {cm (n)} is not invertible. In speech signal processing, the (real) cepstrum has been used to separate and thus to estimate the spectral content of the speech from the pitch frequency of the speech. The complex cepstrum is used in practice to separate signals that are convolved. The process of separating two convolved signals is called deconvolution and the use of the complex cepstrum to perform the separation is called homomorphic deconvolution.

2.8

The Fourier Transform of Signals with Poles on the Unit Circle

As was shown in Section 2.6, the Fourier transform of a sequence x(n) can be determined by evaluating its z-transform X(z) on the unit circle, provided that the unit circle lies within the region of convergence of X(z). Otherwise, the Fourier transform does not exist. There are some aperiodic sequences that are neither absolutely summable nor square summable. Hence their Fourier transforms do not exist. One such sequence is the unit step sequence, which has the z-transform X(z) =

1 1 − z−1

Another such sequence is the causal sinusoidal signal sequence x(n) = (cos ω0 n)u(n). This sequence has the z-transform X(z) =

268

1 − z−1 cos ω0 1 − 2z−1 cos ω0 + z−2

Frequency Analysis of Signals

Note that both of these sequences have poles on the unit circle. For sequences such as these two examples, it is sometimes useful to extend the Fourier transform representation. This can be accomplished, in a mathematically rigorous way, by allowing the Fourier transform to contain impulses at certain frequencies corresponding to the location of the poles of X(z) that lie on the unit circle. The impulses are functions of the continuous frequency variable ω and have infinite amplitude, zero width, and unit area. An impulse can be viewed as the limiting form of a rectangular pulse of height 1/a and width a , in the limit as a → 0. Thus, by allowing impulses in the spectrum of a signal, it is possible to extend the Fourier transform representation to some signal sequences that are neither absolutely summable nor square summable. The following example illustrates the extension of the Fourier transform representation for three sequences. EXAMPLE 2.5 Determine the Fourier transform of the following signals. (a) x1 (n) = u(n) (b) x2 (n) = (−1)n u(n) (c) x3 (n) = (cos ω0 n)u(n) by evaluating their z-transforms on the unit circle. Solution. (a) From common z-transform pairs we find that X1 (z) =

1 z = , 1 − z−1 z−1

ROC: |z| > 1

X1 (z) has a pole, p1 = 1, on the unit circle, but converges for |z| > 1. If we evaluate X1 (z) on the unit circle, except at z = 1, we obtain X1 (ω) =

1 ej ω/2 = ej (ω−π/2) , 2j sin(ω/2) 2 sin(ω/2)

ω = 2πk,

k = 0, 1, . . .

At ω = 0 and multiples of 2π , X1 (ω) contains impulses of area π . Hence the presence of a pole at z = 1 (i.e., at ω = 0) creates a problem only when we want to compute |X1 (ω)| at ω = 0,because |X1 (ω)| → ∞ as ω → 0. For any other value of ω, X1 (ω) is finite (i.e., well behaved). Although at first glance one might expect the signal to have zero-frequency components at all frequencies except at ω = 0, this is not the case. This happens because the signal x1 (n) is not a constant for all −∞ < n < ∞. Instead, it is turned on at n = 0. This abrupt jump creates all frequency components existing in the range 0 < ω ≤ π . Generally, all signals which start at a finite time have nonzero-frequency components everywhere in the frequency axis from zero up to the folding frequency. (b) Given that the z-transform of a n u(n) with a = −1 reduces to X2 (z) =

1 z = , 1 + z−1 z+1

ROC: |z| > 1

269

Frequency Analysis of Signals

which has a pole at z = −1 = ej π . The Fourier transform evaluated at frequencies other than ω = π and multiples of 2π is

X2 (ω) =

ej ω/2 , 2 cos(ω/2)

1 ω = 2π(k + ), 2

k = 0, 1, . . .

In this case the impulse occurs at ω = π + 2πk . Hence the magnitude is

|X2 (ω)| =

1 , 2| cos(ω/2)|

and the phase is ⭿X2 (ω) =

ω = 2πk + π,

   ω2 ,   ω + π, 2

k = 0, 1, . . .

ω ≥0 2 ω if cos < 0 2

if cos

Note that due to the presence of the pole at a = −1 (i.e., at frequency ω = π ), the magnitude of the Fourier transform becomes infinite. Now |X(ω)| → ∞ as ω → π . We observe that (−1)n u(n) = (cos πn)u(n), which is the fastest possible oscillating signal in discrete time. (c) From the discussion above, it follows that X3 (ω) is infinite at the frequency component ω = ω0 . Indeed, we find that

z

x3 (n) = (cos ω0 n)u(n) ←→ X3 (z) =

1 − z−1 cos ω0 , 1 − 2z−1 cos ω0 + z−2

ROC: |z| > 1

The Fourier transform is

X3 (ω) =

1 − e−j ω cos ω0 , (1 − − ej (ω+ω0 ) ) e−j (ω−ω0 ) )(1

ω = ±ω0 + 2πk,

k = 0, 1, . . .

The magnitude of X3 (ω) is given by

|X3 (ω)| =

|1 − e−j ω cos ω0 | , |1 − − e−j (ω+ω0 )| e−j (ω−ω0 )||1

ω = ±ω0 + 2πk,

k = 0, 1, . . .

Now if ω = −ω0 or ω = ω0 , |X3 (ω)| becomes infinite. For all other frequencies, the Fourier transform is well behaved.

270

Frequency Analysis of Signals

Xa(F)

0

X(ω)

F

−π

0

ω

π

(a) Xa(F)

0

X(ω)

F

−π

0

ω

π

(b) Xa(F)

0

X(ω)

F

−π

0

π

ω

(c)

Figure 2.10 (a) Low-frequency, (b) high-frequency, and (c) medium-frequency signals.

2.9

Frequency-Domain Classification of Signals: The Concept of Bandwidth

Just as we have classified signals according to their time-domain characteristics, it is also desirable to classify signals according to their frequency-domain characteristics. It is common practice to classify signals in rather broad terms according to their frequency content. In particular, if a power signal (or energy signal) has its power density spectrum (or its energy density spectrum) concentrated about zero frequency, such a signal is called a low-frequency signal. Figure 2.10(a) illustrates the spectral characteristics of such a signal. On the other hand, if the signal power density spectrum (or the energy density spectrum) is concentrated at high frequencies, the signal is called a high-frequency signal. Such a signal spectrum is illustrated in Fig 2.10(b). A signal having a power density spectrum (or an energy density spectrum) concentrated somewhere in the broad frequency range between low frequencies and high frequencies

271

Continuoustime

Frequency Analysis of Signals

Aperiodic signals

Periodic signals

X(F)

ck

−B

F

B

0

−MF0

Discretetime

−ω0

0

MF0

kF0

ck

X(ω)

−π

0

ω0

π

ω

−π

−M2π N

0

M2π N

π

ω

Figure 2.11 Some examples of bandlimited signals.

is called a medium-frequency signal or a bandpass signal. Figure 2.10(c) illustrates such a signal spectrum. In addition to this relatively broad frequency-domain classification of signals, it is often desirable to express quantitatively the range of frequencies over which the power or energy density spectrum is concentrated. This quantitative measure is called the bandwidth of a signal. For example, suppose that a continuous-time signal has 95% of its power (or energy) density spectrum concentrated in the frequency range F1 ≤ F ≤ F2 . Then the 95% bandwidth of the signal is F2 − F1 . In a similar manner, we may define the 75% or 90% or 99% bandwidth of the signal. In the case of a bandpass signal, the term narrowband is used to describe the signal if its bandwidth F2 − F1 is much smaller (say, by a factor of 10 or more) than the median frequency (F2 + F1 )/2. Otherwise, the signal is called wideband. We shall say that a signal is bandlimited if its spectrum is zero outside the frequency range |F | ≥ B . For example, a continuous-time finite-energy signal x(t) is bandlimited if its Fourier transform X(F ) = 0 for |F | > B . A discrete-time finite-energy signal x(n) is said to be (periodically) bandlimited if |X(ω)| = 0,

for ω0 < |ω| < π

Similarly, a periodic continuous-time signal xp (t) is periodically bandlimited if its Fourier coefficients ck = 0 for |k| > M , where M is some positive integer. A periodic discrete-time signal with fundamental period N is periodically bandlimited if the Fourier coefficients ck = 0 for k0 < |k| < N . Figure 2.11 illustrates the four types of bandlimited signals. By exploiting the duality between the frequency domain and the time domain, we can provide similar means for characterizing signals in the time domain. In particular, a signal x(t) will be called time-limited if x(t) = 0,

272

|t| > τ

Frequency Analysis of Signals

If the signal is periodic with period Tp , it will be called periodically time-limited if xp (t) = 0,

τ < |t| < Tp /2

If we have a discrete-time signal x(n) of finite duration, that is, x(n) = 0,

|n| > N

it is also called time-limited. When the signal is periodic with fundamental period N , it is said to be periodically time-limited if x(n) = 0,

n0 < |n| < N

We state, without proof, that no signal can be time-limited and bandlimited simultaneously. Furthermore, a reciprocal relationship exists between the time duration and the frequency duration of a signal. To elaborate, if we have a short-duration rectangular pulse in the time domain, its spectrum has a width that is inversely proportional to the duration of the time-domain pulse. The narrower the pulse becomes in the time domain, the larger the bandwidth of the signal becomes. Consequently, the product of the time duration and the bandwidth of a signal cannot be made arbitrarily small. A short-duration signal has a large bandwidth and a small bandwidth signal has a long duration. Thus, for any signal, the time–bandwidth product is fixed and cannot be made arbitrarily small. Finally, we note that we have discussed frequency analysis methods for periodic and aperiodic signals with finite energy. However, there is a family of deterministic aperiodic signals with finite power. These signals consist of a linear superposition of complex exponentials with nonharmonically related frequencies, that is, x(n) =

M 

A k e j ωk n

k=1

where ω1 , ω2 , . . . , ωM are nonharmonically related. These signals have discrete spectra but the distances among the lines are nonharmonically related. Signals with discrete nonharmonic spectra are sometimes called quasi-periodic.

2.10

The Frequency Ranges of Some Natural Signals

The frequency analysis tools that we have developed in this chapter are usually applied to a variety of signals that are encountered in practice (e.g., seismic, biological, and electromagnetic signals). In general, the frequency analysis is performed for the purpose of extracting information from the observed signal. For example, in the case of biological signals, such as an ECG signal, the analytical tools are used to extract information relevant for diagnostic purposes. In the case of seismic signals, we may be interested in detecting the presence of a nuclear explosion or in determining the characteristics and location of an earthquake. An electromagnetic signal, such as a radar signal reflected from an airplane, contains information on the position of the

273

Frequency Analysis of Signals

plane and its radial velocity. These parameters can be estimated from observation of the received radar signal. In processing any signal for the purpose of measuring parameters or extracting other types of information, one must know approximately the range of frequencies contained by the signal. For reference, Tables 1, 2, and 3 give approximate limits in the frequency domain for biological, seismic, and electromagnetic signals.

3

Frequency-Domain and Time-Domain Signal Properties In the previous sections of the chapter we have introduced several methods for the frequency analysis of signals. Several methods were necessary to accommodate the different types of signals. To summarize, the following frequency analysis tools have been introduced: 1. The Fourier series for continuous-time periodic signals. 2. The Fourier transform for continuous-time aperiodic signals. 3. The Fourier series for discrete-time periodic signals. 4. The Fourier transform for discrete-time aperiodic signals. TABLE 1 Frequency Ranges of Some Biological Signals

Type of Signal

Frequency Range (Hz)

Electroretinogram a Electronystagmogram

0–20 b

0–20

Pneumogram c

0–40

Electrocardiogram (ECG)

0–100

Electroencephalogram (EEG) Electromyogram d Sphygmomanogram e Speech

0–100 10–200 0–200 100–4000

aA

graphic recording of retina characteristics. graphic recording of involuntary movement of the eyes. c A graphic recording of respiratory activity. d A graphic recording of muscular action, such as muscular contraction. e A recording of blood pressure. bA

TABLE 2 Frequency Ranges of Some Seismic Signals

Type of Signal Wind noise Seismic exploration signals Earthquake and nuclear explosion signals Seismic noise

274

Frequency Range (Hz) 100–1000 10–100 0.01–10 0.1–1

Frequency Analysis of Signals

TABLE 3 Frequency Ranges of Electromagnetic Signals

Type of Signal Radio broadcast Shortwave radio signals

Wavelength (m) 4

3 × 104 –3 × 106

−2

3 × 106 –3 × 1010

10 –10 2

10 –10

Frequency Range (Hz)

2

Radar, satellite communications, space communications, common-carrier microwave Infrared Visible light

1–10 −2

3 × 108 –3 × 1010

10 −3 –10 −6 3.9 × 10

−7

–8.1 × 10

3 × 1011 –3 × 1014 −7

3.7 × 1014 –7.7 × 1014

Ultraviolet

10 −7 –10 −8

3 × 1015 –3 × 1016

Gamma rays and X rays

10 −9 –10 −10

3 × 1017 –3 × 1018

Figure 3.1 summarizes the analysis and synthesis formulas for these types of signals. As we have already indicated several times, there are two time-domain characteristics that determine the type of signal spectrum we obtain. These are whether the time variable is continuous or discrete, and whether the signal is periodic or aperiodic. Let us briefly summarize the results of the previous sections. Continuous-time signals have aperiodic spectra A close inspection of the Fourier series and Fourier transform analysis formulas for continuous-time signals does not reveal any kind of periodicity in the spectral domain. This lack of periodicity is a consequence of the fact that the complex exponential exp(j 2π F t) is a function of the continuous variable t, and hence it is not periodic in F . Thus the frequency range of continuous-time signals extends from F = 0 to F = ∞. Discrete-time signals have periodic spectra Indeed, both the Fourier series and the Fourier transform for discrete-time signals are periodic with period ω = 2π . As a result of this periodicity, the frequency range of discrete-time signals is finite and extends from ω = −π to ω = π radians, where ω = π corresponds to the highest possible rate of oscillation. Periodic signals have discrete spectra As we have observed, periodic signals are described by means of Fourier series. The Fourier series coefficients provide the “lines” that constitute the discrete spectrum. The line spacing F or f is equal to the inverse of the period Tp or N , respectively, in the time domain. That is, F = 1/Tp for continuous-time periodic signals and f = 1/N for discrete-time signals. Aperiodic finite energy signals have continuous spectra This property is a direct consequence of the fact that both X(F ) and X(ω) are functions of exp(j 2π F t) and exp(j ωn), respectively, which are continuous functions of the variables F and ω. The continuity in frequency is necessary to break the harmony and thus create aperiodic signals.

275

276 Continuous-time signals Time-domain

Frequency-domain

xa(t)

x(n)

… −Tp

ck =

0

Tp

F0 =

Σ k=−

xa(t)

1 Tp

ck =

0

N−1

Discrete and periodic

Σ

X(ω)



… −3 −2 −1 0 1 2

Σ

Figure 3.1 Summary of analysis and synthesis formulas.

… n

… −2π

−π

0

π



x(n)e−jωn

n=−

Xa(F)e j2πFt dF

Continuous and aperiodic

ck e j(2π/N)kn

Discrete and periodic

x(n)

F

k

N

1 N−1 x(n)e−j(2π/N)kn N nΣ =0

X(ω) =



… −N

k=0

0

xa(t) = ∫

… n

N

x(n) =

Xa(F)

t

0

ck e j2πkF0t

xa(t)e−j2πFt dt

Continuous and aperiodic



−N

Discrete and aperiodic

Continuous and periodic



F

0

xa(t) =

Xa(F) = ∫

ck

… t

1 x (t)e−j2πkF0t dt Tp ∫Tp a

0

Frequency-domain

x(n) = Discrete and aperiodic

1 X(ω)e jωn dω 2π ∫2π

Continuous and periodic

ω

Frequency Analysis of Signals

Periodic signals Fourier series

Time-domain

ck



Aperiodic signals Fourier transforms

Discrete-time signals

Frequency Analysis of Signals

In summary, we can conclude that periodicity with “period” α in one domain automatically implies discretization with “spacing” of 1/α in the other domain, and vice versa. If we keep in mind that “period” in the frequency domain means the frequency range, “spacing” in the time domain is the sampling period T , line spacing in the frequency domain is F , then α = Tp implies that 1/α = 1/Tp = F , α = N implies that f = 1/N , and α = Fs implies that T = 1/Fs . These time-frequency dualities are apparent from observation of Fig 3.1. We stress, however, that the illustrations used in this figure do not correspond to any actual transform pairs. Thus any comparison among them should be avoided. A careful inspection of Fig 3.1 also reveals some mathematical symmetries and dualities among the several frequency analysis relationships. In particular, we observe that there are dualities between the following analysis and synthesis equations: 1. The analysis and synthesis equations of the continuous-time Fourier transform. 2. The analysis and synthesis equations of the discrete-time Fourier series. 3. The analysis equation of the continuous-time Fourier series and the synthesis equation of the discrete-time Fourier transform. 4. The analysis equation of the discrete-time Fourier transform and the synthesis equation of the continuous-time Fourier series. Note that all dual relations differ only in the sign of the exponent of the corresponding complex exponential. It is interesting to note that this change in sign can be thought of either as a folding of the signal or a folding of the spectrum, since e−j 2πF t = ej 2π(−F )t = ej 2πF (−t) If we turn our attention now to the spectral density of signals, we recall that we have used the term energy density spectrum for characterizing finite-energy aperiodic signals and the term power density spectrum for periodic signals. This terminology is consistent with the fact that periodic signals are power signals and aperiodic signals with finite energy are energy signals.

4

Properties of the Fourier Transform for Discrete-Time Signals The Fourier transform for aperiodic finite-energy discrete-time signals described in the preceding section possesses a number of properties that are very useful in reducing the complexity of frequency analysis problems in many practical applications. In this section we develop the important properties of the Fourier transform. Similar properties hold for the Fourier transform of aperiodic finite-energy continuous-time signals.

277

Frequency Analysis of Signals

For convenience, we adopt the notation X(ω) ≡ F {x(n)} =

∞ 

x(n)e−j ωn

(4.1)

X(ω)ej ωn dω

(4.2)

n=−∞

for the direct transform (analysis equation) and x(n) ≡ F −1 {X(ω)} =

1 2π

 2π

for the inverse transform (synthesis equation). We also refer to x(n) and X(ω) as a Fourier transform pair and denote this relationship with the notation F

x(n) ←→ X(ω)

(4.3)

Recall that X(ω) is periodic with period 2π . Consequently, any interval of length 2π is sufficient for the specification of the spectrum. Usually, we plot the spectrum in the fundamental interval [−π, π]. We emphasize that all the spectral information contained in the fundamental interval is necessary for the complete description or characterization of the signal. For this reason, the range of integration in( 4.2) is always 2π , independent of the specific characteristics of the signal within the fundamental interval.

4.1

Symmetry Properties of the Fourier Transform

When a signal satisfies some symmetry properties in the time domain, these properties impose some symmetry conditions on its Fourier transform. Exploitation of any symmetry characteristics leads to simpler formulas for both the direct and inverse Fourier transform. A discussion of various symmetry properties and the implications of these properties in the frequency domain is given here. Suppose that both the signal x(n) and its transform X(ω) are complex-valued functions. Then they can be expressed in rectangular form as x(n) = xR (n) + j xI (n)

(4.4)

X(ω) = XR (ω) + j XI (ω)

(4.5)

By substituting (4.4) and e −j ω = cos ω − j sin ω into (4.1) and separating the real and imaginary parts, we obtain XR (ω) =

∞ 

[xR (n) cos ωn + xI (n) sin ωn]

(4.6)

n=−∞

XI (ω) = −

∞  n=−∞

278

[xR (n) sin ωn − xI (n) cos ωn]

(4.7)

Frequency Analysis of Signals

In a similar manner, by substituting (4.5) and e j ω = cos ω + j sin ω into (4.2), we obtain  1 [XR (ω) cos ωn − XI (ω) sin ωn] dω (4.8) xR (n) = 2π 2π  1 [XR (ω) sin ωn + XI (ω) cos ωn] dω xI (n) = (4.9) 2π 2π Now, let us investigate some special cases. Real signals. If x(n) is real, then xR (n) = x(n) and xI (n) = 0. Hence (4.6) and (4.7) reduce to ∞  (4.10) x(n) cos ωn XR (ω) = n=−∞

and

∞ 

XI (ω) = −

x(n) sin ωn

(4.11)

n=−∞

Since cos(−ωn) = cos ωn and sin(−ωn) = − sin ωn, it follows from (4.10) and (4.11) that XR (−ω) = XR (ω) ,

(even)

(4.12)

XI (−ω) = −XI (ω),

(odd)

(4.13)

If we combine (4.12) and (4.13) into a single equation, we have X∗ (ω) = X(−ω)

(4.14)

In this case we say that the spectrum of a real signal has Hermitian symmetry. With the aid of Fig 4.1, we observe that the magnitude and phase spectra for real signals are |X(ω)| =



XR2 (ω) + XI2 (ω)

(4.15)

XI (ω) XR (ω)

(4.16)

⭿X|ω| = tan−1

As a consequence of (4.12) and (4.13), the magnitude and phase spectra also possess the symmetry properties |X(ω)| = |X(−ω)| , ⭿X(−ω) = −⭿X(ω),

(even)

(4.17)

(odd)

(4.18)

279

Frequency Analysis of Signals

Imaginary axis

X(ω)

XI(ω)

X(

ω)

Figure 4.1

X(ω)

Magnitude and phase functions.

0

Real axis

XR(ω)

In the case of the inverse transform of a real-valued signal [i.e., x(n) = xR (n)], (4.8) implies that x(n) =

1 2π

 [XR (ω) cos ωn − XI (ω) sin ωn] dω

(4.19)



Since both products XR (ω) cos ωn and XI (ω) sin ωn are even functions of ω, we have 1 x(n) = π



π

[XR (ω) cos ωn − XI (ω) sin ωn] dω

(4.20)

0

If x(n) is real and even [i.e., x(−n) = x(n)], then x(n) cos ωn is even and x(n) sin ωn is odd. Hence, from (4.10), (4.11), and (4.20) we obtain

Real and even signals.

XR (ω) = x(0) + 2

∞ 

x(n) cos ωn,

(even)

(4.21)

n=1

XI (ω) = 0 x(n) =

1 π

(4.22) 

π

XR (ω) cos ωn dω

(4.23)

0

Thus real and even signals possess real-valued spectra, which, in addition, are even functions of the frequency variable ω. If x(n) is real and odd [i.e., x(−n) = −x(n)], then x(n) cos ωn is odd and x(n) sin ωn is even. Consequently, (4.10), (4.11) and (4.20) imply that

Real and odd signals.

XR (ω) = 0

(4.24)

XI (ω) = −2

∞ 

x(n) sin ωn,

(odd)

(4.25)

n=1

x(n) = −

280

1 π



π

XI (ω) sin ωn dω 0

(4.26)

Frequency Analysis of Signals

Thus real-valued odd signals possess purely imaginary-valued spectral characteristics, which, in addition, are odd functions of the frequency variable ω. Purely imaginary signals. In this case xR (n) = 0 and x(n) = j xI (n). Thus (4.6),

(4.7), and (4.9) reduce to

XR (ω) =

∞ 

xI (n) sin ωn,

(odd)

(4.27)

xI (n) cos ωn,

(even)

(4.28)

n=−∞

XI (ω) =

∞  n=−∞

1 xI (n) = π



π

[XR (ω) sin ωn + XI (ω) cos ωn] dω

(4.29)

0

If xI (n) is odd [i.e., xI (−n) = −xI (n)], then

XR (ω) = 2

∞ 

xI (n) sin ωn,

(odd)

(4.30)

n=1

XI (ω) = 0 xI (n) =

1 π

(4.31) 

π

(4.32)

XR (ω) sin ωn dω 0

Similarly, if xI (n) is even [i.e., xI (−n) = xI (n)], we have XR (ω) = 0

(4.33)

XI (ω) = xI (0) + 2 1 xI (n) = π



∞ 

xI (n) cos ωn,

(even)

(4.34)

n=1 π

XI (ω) cos ωn dω

(4.35)

0

An arbitrary, possibly complex-valued signal x(n) can be decomposed as x(n) = xR (n) + j xI (n) = xRe (n) + xRo (n) + j [xIe (n) + xIo (n)] = xe (n) + xo (n) (4.36)

281

Frequency Analysis of Signals

where, by definition,

xe (n) = xRe (n) + j xIe (n) =

1 [x(n) + x ∗ (−n)] 2

xo (n) = xRo (n) + j xIo (n) =

1 [x(n) − x ∗ (−n)] 2

The superscripts e and o denote the even and odd signal components, respectively. We note that xe (n) = xe (−n) and xo (−n) = −xo (n). From (4.36) and the Fourier transform properties established above, we obtain the following relationships:

x(n) = [xRe(n) + jx Ie(n)] + [xRo(n) + jxIo(n)] = xe(n) + xo(n) X(ω) = [XRe(ω) + jXIe(ω)] + [X oR (ω) − jXIo(ω)] = Xe(ω) + Xo(ω)

(4.37)

These symmetry properties of the Fourier transform are summarized in Table 4 and in Fig 4.2. They are often used to simplify Fourier transform calculations in practice.

TABLE 4 Symmetry Properties of the Discrete-Time Fourier Transform

Sequence x(n) x ∗ (n) x ∗ (−n) xR (n) j xI (n) xe (n) = 21 [x(n) + x ∗ (−n)] xo (n) = 21 [x(n) − x ∗ (−n)]

DTFT X(ω) X ∗ (−ω) X ∗ (ω) Xe (ω) = 21 [X(ω) + X ∗ (−ω)] Xo (ω) = 21 [X(ω) − X ∗ (−ω)] XR (ω) j XI (ω) Real Signals

Any real signal x(n)

xe (n) = 21 [x(n) + x(−n)] (real and even) xo (n) = 21 [x(n) − x(−n)] (real and odd)

282

X(ω) = X ∗ (−ω) XR (ω) = XR (−ω) XI (ω) = −XI (−ω) |X(ω)| = |X(−ω)| ⭿X(ω) = −⭿X(−ω) XR (ω) (real and even) j XI (ω) (imaginary and odd)

Frequency Analysis of Signals

Time domain

Frequency domain

Even

Even

Odd

Odd

Real

Real

Signal

Fourier Transform

Odd

Odd

Even

Even

Imaginary

Imaginary

Figure 4.2 Summary of symmetry properties for the Fourier transform.

EXAMPLE 4.1 Determine and sketch XR (ω), XI (ω), |X(ω)|, and ⭿X(ω) for the Fourier transform X(ω) =

1 , 1 − ae−j ω

−1 < a < 1

(4.38)

Solution. By multiplying both the numerator and denominator of (4.38) by the complex conjugate of the denominator, we obtain X(ω) =

1 − aej ω 1 − a cos ω − j a sin ω = (1 − ae−j ω )(1 − aej ω ) 1 − 2a cos ω + a 2

This expression can be subdivided into real and imaginary parts. Thus we obtain XR (ω) =

1 − a cos ω 1 − 2a cos ω + a 2

XI (ω) = −

a sin ω 1 − 2a cos ω + a 2

Substitution of the last two equations into (4.15) and (4.16) yields the magnitude and phase spectra as 1 |X(ω)| = √ (4.39) 1 − 2a cos ω + a 2 and ⭿X(ω) = − tan−1

a sin ω 1 − a cos ω

(4.40)

Figures 4.3 and 4.4 show the graphical representation of these spectra for a = 0.8. The reader can easily verify that as expected, all symmetry properties for the spectra of real signals apply to this case.

283

Frequency Analysis of Signals

6 5 X R (ω)

4 3 2 1 0

π

0. 5 π

0

0.5 π

π

0. 5 π

0

0.5 π

π

0. 5 π

0

0.5 π

π

0. 5 π

0

0.5 π

ω

π

3 2 X1(ω)

1 0 1 2

Figure 4.3

Graph of XR (ω) and XI (ω) for the transform in Example 4.1.

3

ω

π

6 5 |X (ω)|

4 3 2 1 0

ω

π

0.4π

∠X( ω)

0.2π 0 0.2π Figure 4.4

Magnitude and phase spectra of the transform in Example 4.1.

284

0.4π

ω

π

Frequency Analysis of Signals

EXAMPLE 4.2 Determine the Fourier transform of the signal  x(n) = Solution. obtain

−M ≤ n ≤ M elsewhere

A, 0,

(4.41)

Clearly, x(−n) = x(n). Thus x(n) is a real and even signal. From (4.21) we  X(ω) = XR (ω) = A 1 + 2

M 

 cos ωn

n=1

If we use the identity given in Problem 13, we obtain the simpler form X(ω) = A

sin(M + 21 )ω sin(ω/2)

Since X(ω) is real, the magnitude and phase spectra are given by    sin(M + 1 )ω    2 |X(ω)| = A   sin(ω/2)  and

 ⭿X(ω) =

0, π,

if X(ω) > 0 if X(ω) < 0

(4.42)

(4.43)

Figure 4.5 shows the graphs for X(ω).

4.2

Fourier Transform Theorems and Properties

In this section we introduce several Fourier transform theorems and illustrate their use in practice by examples. Linearity.

If F

x1 (n) ←→ X1 (ω) and F

x2 (n) ←→ X2 (ω) then F

a1 x1 (n) + a2 x2 (n) ←→ a1 X1 (ω) + a2 X2 (ω)

(4.44)

Simply stated, the Fourier transformation, viewed as an operation on a signal x(n), is a linear transformation. Thus the Fourier transform of a linear combination of two or more signals is equal to the same linear combination of the Fourier transforms of the individual signals. This property is easily proved by using (4.1). The linearity property makes the Fourier transform suitable for the study of linear systems.

285

Frequency Analysis of Signals

x(n)

n

−M 0 M X(ω) 4 2 −2π

−π

π

−2



ω

|X(ω)|

4 2

−2π

−π

0

π



π



ω

X(ω) π

−2π

−π

ω

−π

Figure 4.5 Spectral characteristics of rectangular pulse in

Example 4.2.

EXAMPLE 4.3 Determine the Fourier transform of the signal x(n) = a |n| , Solution.

−1 < a < 1

First, we observe that x(n) can be expressed as x(n) = x1 (n) + x2 (n)

where

 x1 (n) =

286

an, 0,

n≥0 n 1) or attenuation (|H (ω)| < 1) imparted by the system on the input sinusoid. The phase (ω) determines the amount of phase shift imparted by the system on the input sinusoid. Consequently, by knowing H (ω), we are able to determine the response of the system to any sinusoidal input signal. Since H (ω) specifies the response of the system in the frequency domain, it is called the frequency response of the system. Correspondingly, |H (ω)| is called the magnitude response and (ω) is called the phase response of the system. If the input to the system consists of more than one sinusoid, the superposition property of the linear system can be used to determine the response. The following examples illustrate the use of the superposition property. EXAMPLE 1.3 Determine the response of the system in Example 1.1 to the input signal x(n) = 10 − 5 sin Solution.

π n + 20 cos πn, 2

−∞ < n < ∞

The frequency response of the system is given in (1.7) as H (ω) =

1 1−

1 −j ω e 2

The first term in the input signal is a fixed signal component corresponding to ω = 0. Thus H (0) =

314

1 1−

1 2

=2

Frequency-Domain Analysis of LTI Systems

The second term in x(n) has a frequency π/2. At this frequency the frequency response of the system is π  2 ◦ = √ e−j 26.6 H 2 5 Finally, the third term in x(n) has a frequency ω = π . At this frequency H (π) =

2 3

Hence the response of the system to x(n) is  40 π 10 y(n) = 20 − √ sin n − 26.6◦ + cos πn, 2 3 5

−∞ < n < ∞

EXAMPLE 1.4 A linear time-invariant system is described by the following difference equation: y(n) = ay(n − 1) + bx(n),

0 21 , the inverse transform yields 1 hI (n) = ( )n u(n) 2 which is the impulse response of a causal and stable system. On the other hand, if the ROC is assumed to be |z| < 21 , the inverse system has an impulse response hI (n) = −

 n 1 u(−n − 1) 2

In this case the inverse system is anticausal and unstable.

360

Frequency-Domain Analysis of LTI Systems

ROC

ROC

z-plane 0

1 2

1 2

(a)

Figure 5.2

z-plane

(b)

Two possible regions of convergence for H (z) = z/(z − 21 ).

We observe that (5.3) cannot be solved uniquely by using (5.6) unless we specify the region of convergence for the system function of the inverse system. In some practical applications the impulse response h(n) does not possess a ztransform that can be expressed in closed form. As an alternative we may solve (5.3) directly using a digital computer. Since (5.3) does not, in general, possess a unique solution, we assume that the system and its inverse are causal. Then (5.3) simplifies to the equation n 

h(k)hI (n − k) = δ(n)

(5.7)

k=0

By assumption, hI (n) = 0 for n < 0. For n = 0 we obtain hI (0) = 1/ h(0)

(5.8)

The values of hI (n) for n ≥ 1 can be obtained recursively from the equation hI (n) =

n  h(n)hI (n − k) k=1

h(0)

,

n≥1

(5.9)

This recursive relation can easily be programmed on a digital computer. There are two problems associated with (5.9). First, the method does not work if h(0) = 0. However, this problem can easily be remedied by introducing an appropriate delay in the right-hand side of (5.7), that is, by replacing δ(n) by δ(n − m), where m = 1 if h(0) = 0 and h(1) = 0, and so on. Second, the recursion in (5.9) gives rise to round-off errors which grow with n and, as a result, the numerical accuracy of h(n) deteriorates for large n.

361

Frequency-Domain Analysis of LTI Systems

EXAMPLE 5.3 Determine the causal inverse of the FIR system with impulse response h(n) = δ(n) − αδ(n − 1) Since h(0) = 1, h(1) = −α , and h(n) = 0 for n ≥ α , we have hI (0) = 1/ h(0) = 1 and hI (n) = αhI (n − 1),

n≥1

Consequently, hI (1) = α,

hI (2) = α 2 ,

...,

hI (n) = α n

which corresponds to a causal IIR system as expected.

5.2

Minimum-Phase, Maximum-Phase, and Mixed-Phase Systems

The invertibility of a linear time-invariant system is intimately related to the characteristics of the phase spectral function of the system. To illustrate this point, let us consider two FIR systems, characterized by the system functions 1 1 H1 (z) = 1 + z−1 = z−1 (z + ) 2 2

(5.10)

1 1 + z−1 = z−1 ( z + 1) 2 2

(5.11)

H2 (z) =

The system in (5.10) has a zero at z = − 21 and an impulse response h(0) = 1, h(1) = 1/2. The system in (5.11) has a zero at z = −2 and an impulse response h(0) = 1/2, h(1) = 1, which is the reverse of the system in (5.10). This is due to the reciprocal relationship between the zeros of H1 (z) and H2 (z). In the frequency domain, the two systems are characterized by their frequency response functions, which can be expressed as  |H1 (ω)| = |H2 (ω)| =

5 + cos ω 4

(5.12)

and 1 (ω) = −ω + tan−1 2 (ω) = −ω + tan−1

362

sin ω 1 2

+ cos ω

sin ω 2 + cos ω

(5.13)

(5.14)

Frequency-Domain Analysis of LTI Systems

θ1(ω)

θ2(ω)

π

π

ω

ω

−π

−π (a)

Figure 5.3

(b)

Phase response characteristics for the systems in (5.10). and (5.11).

The magnitude characteristics for the two systems are identical because the zeros of H1 (z) and H2 (z) are reciprocals. The graphs of 1 (ω) and 2 (ω) are illustrated in Fig. 5.3. We observe that the phase characteristic 1 (ω) for the first system begins at zero phase at the frequency ω = 0 and terminates at zero phase at the frequency ω = π . Hence the net phase change, 1 (π ) − 1 (0), is zero. On the other hand, the phase characteristic for the system with the zero outside the unit circle undergoes a net phase change 2 (π ) − 2 (0) = π radians. As a consequence of these different phase characteristics, we call the first system a minimum-phase system and the second system a maximum-phase system. These definitions are easily extended to an FIR system of arbitrary length. To be specific, an FIR system of length M + 1 has M zeros. Its frequency response can be expressed as H (ω) = b0 (1 − z1 e−j ω )(1 − z2 e−j ω ) · · · (1 − zM e−j ω )

(5.15)

where {zi } denote the zeros and b0 is an arbitrary constant. When all the zeros are inside the unit circle, each term in the product of (5.15), corresponding to a realvalued zero, will undergo a net phase change of zero between ω = 0 and ω = π . Also, each pair of complex-conjugate factors in H (ω) will undergo a net phase change of zero. Therefore, ⭿H (π ) − ⭿H (0) = 0 (5.16) and hence the system is called a minimum-phase system. On the other hand, when all the zeros are outside the unit circle, a real-valued zero will contribute a net phase change of π radians as the frequency varies from ω = 0 to ω = π , and each pair of complex-conjugate zeros will contribute a net phase change of 2π radians over the same range of ω. Therefore, ⭿H (π ) − ⭿H (0) = Mπ

(5.17)

which is the largest possible phase change for an FIR system with M zeros. Hence the system is called maximum phase. It follows from the discussion above that ⭿Hmax (π ) ≥ ⭿Hmin (π )

(5.18)

363

Frequency-Domain Analysis of LTI Systems

If the FIR system with M zeros has some of its zeros inside the unit circle and the remaining zeros outside the unit circle, it is called a mixed-phase system or a nonminimum-phase system. Since the derivative of the phase characteristic of the system is a measure of the time delay that signal frequency components undergo in passing through the system, a minimum-phase characteristic implies a minimum delay function, while a maximum-phase characteristic implies that the delay characteristic is also maximum. Now suppose that we have an FIR system with real coefficients. Then the magnitude square value of its frequency response is |H (ω)|2 = H (z)H (z−1 )|z=ej ω

(5.19)

This relationship implies that if we replace a zero zk of the system by its inverse 1/zk , the magnitude characteristic of the system does not change. Thus if we reflect a zero zk that is inside the unit circle into a zero 1/zk outside the unit circle, we see that the magnitude characteristic of the frequency response is invariant to such a change. It is apparent from this discussion that if |H (ω)|2 is the magnitude square frequency response of an FIR system having M zeros, there are 2M possible configurations for the M zeros, of which some are inside the unit circle and the remaining are outside the unit circle. Clearly, one configuration has all the zeros inside the unit circle, which corresponds to the minimum-phase system. A second configuration has all the zeros outside the unit circle, which corresponds to the maximum-phase system. The remaining 2M − 2 configurations correspond to mixed-phase systems. However, not all 2M − 2 mixed-phase configurations necessarily correspond to FIR systems with real-valued coefficients. Specifically, any pair of complex-conjugate zeros results in only two possible configurations, whereas a pair of real-valued zeros yields four possible configurations. EXAMPLE 5.4 Determine the zeros for the following FIR systems and indicate whether the system is minimum phase, maximum phase, or mixed phase. H1 (z) = 6 + z−1 − z−2 H2 (z) = 1 − z−1 − 6z−2 5 3 H3 (z) = 1 − z−1 − z−2 2 2 5 2 H4 (z) = 1 + z−1 − z−2 3 3 Solution.

By factoring the system functions we find the zeros for the four systems are 1 1 H1 (z) −→ z1,2 = − , −→ minimum phase 2 3 H2 (z) −→ z1,2 = −2, 3 −→ maximum phase

364

Frequency-Domain Analysis of LTI Systems

1 H3 (z) −→ z1,2 = − , 3 −→ mixed phase 2 H4 (z) −→ z1,2 = −2,

1 −→ mixed phase 3

Since the zeros of the four systems are reciprocals of one another, it follows that all four systems have identical magnitude frequency response characteristics but different phase characteristics.

The minimum-phase property of FIR systems carries over to IIR systems that have rational system functions. Specifically, an IIR system with system function H (z) =

B(z) A(z)

(5.20)

is called minimum phase if all its poles and zeros are inside the unit circle. For a stable and causal system [all roots of A(z) fall inside the unit circle] the system is called maximum phase if all the zeros are outside the unit circle, and mixed phase if some, but not all, of the zeros are outside the unit circle. This discussion brings us to an important point that should be emphasized. That is, a stable pole–zero system that is minimum phase has a stable inverse which is also minimum phase. The inverse system has the system function H −1 (z) =

A(z) B(z)

(5.21)

Hence the minimum-phase property of H (z) ensures the stability of the inverse system H −1 (z) and the stability of H (z) implies the minimum-phase property of H −1 (z). Mixed-phase systems and maximum-phase systems result in unstable inverse systems. Decomposition of nonminimum-phase pole–zero systems. Any nonminimum-phase pole–zero system can be expressed as H (z) = Hmin (z)Hap (z)

(5.22)

where Hmin (z) is a minimum-phase system and Hap (z) is an all-pass system. We demonstrate the validity of this assertion for the class of causal and stable systems with a rational system function H (z) = B(z)/A(z). In general, if B(z) has one or more roots outside the unit circle, we factor B(z) into the product B1 (z)B2 (z), where B1 (z) has all its roots inside the unit circle and B2 (z) has all its roots outside the unit circle. Then B2 (z−1 ) has all its roots inside the unit circle. We define the minimum-phase system B1 (z)B2 (z−1 ) Hmin (z) = A(z) and the all-pass system Hap (z) =

B2 (z) B2 (z−1 )

Thus H (z) = Hmin (z)Hap (z). Note that Hap (z) is a stable, all-pass, maximum-phase system.

365

Frequency-Domain Analysis of LTI Systems

Group delay of nonminimum-phase system. Based on the decomposition of a nonminimum-phase system given by (5.22), we can express the group delay of H (z) as τg (ω) = τgmin (ω) + τgap (ω)

(5.23)

ap

Since τg (ω) ≥ 0 for 0 ≤ ω ≤ π , it follows that τg (ω) ≥ τgmin (ω), 0 ≤ ω ≤ π . From (5.23) we conclude that among all pole–zero systems having the same magnitude response, the minimum-phase system has the smallest group delay. Partial energy of nonminimum-phase system. The partial energy of a causal system with impulse response h(n) is defined as E(n) =

n 

|h(k)|2

(5.24)

k=0

It can be shown that among all systems having the same magnitude response and the same total energy E(∞), the minimum-phase system has the largest partial energy [i.e., Emin (n) ≥ E(n), where Emin (n) is the partial energy of the minimum-phase system].

5.3

System Identification and Deconvolution

Suppose that we excite an unknown linear time-invariant system with an input sequence x(n) and we observe the output sequence y(n). From the output sequence we wish to determine the impulse response of the unknown system. This is a problem in system identification, which can be solved by deconvolution. Thus we have y(n) = h(n) ∗ x(n) =

∞ 

h(k)x(n − k)

(5.25)

k=−∞

An analytical solution of the deconvolution problem can be obtained by working with the z-transform of (5.25). In the z-transform domain we have Y (z) = H (z)X(z) and hence H (z) =

Y (z) X(z)

(5.26)

X(z) and Y (z) are the z-transforms of the available input signal x(n) and the observed output signal y(n), respectively. This approach is appropriate only when there are closed-form expressions for X(z) and Y (z).

366

Frequency-Domain Analysis of LTI Systems

EXAMPLE 5.5 A causal system produces the output sequence y(n) =

  1, 

7 , 10

0,

n=0 n=1 otherwise

when excited by the input sequence  1,    7 − 10 , x(n) = 1  10 ,   0,

n=0 n=1 n=2 otherwise

Determine its impulse response and its input–output equation. Solution. The system function is easily determined by taking the z-transforms of x(n) and y(n). Thus we have H (z) =

=

7 −1 1 + 10 z Y (z) = 7 −1 1 −2 X(z) 1 − 10 z + 10 z

(1 −

7 −1 1 + 10 z 1 −1 z )(1 − 15 z−1 ) 2

Since the system is causal, its ROC is |z| > 21 . The system is also stable since its poles lie inside the unit circle. The input–output difference equation for the system is y(n) =

1 7 7 y(n − 1) − y(n − 2) + x(n) + x(n − 1) 10 10 10

Its impulse response is determined by performing a partial-fraction expansion of H (z) and inverse transforming the result. This computation yields 1 1 h(n) = [4( )n − 3( )n ]u(n) 2 5

We observe that (5.26) determines the unknown system uniquely if it is known that the system is causal. However, the example above is artificial, since the system response {y(n)} is very likely to be infinite in duration. Consequently, this approach is usually impractical. As an alternative, we can deal directly with the time-domain expression given by (5.25). If the system is causal, we have y(n) =

n 

h(k)x(n − k),

n≥0

k=0

367

Frequency-Domain Analysis of LTI Systems

and hence h(0) =

y(0) x(0) y(n) −

n−1  k=0

h(n) =

(5.27)

h(k)x(n − k)

x(0)

,

n≥1

This recursive solution requires that x(0) = 0. However, we note again that when {h(n)} has infinite duration, this approach may not be practical unless we truncate the recursive solution at same stage [i.e., truncate {h(n)}]. Another method for identifying an unknown system is based on a crosscorrelation technique. Recall that the input–output crosscorrelation function is given as ryx (m) =

∞ 

h(k)rxx (m − k) = h(n) ∗ rxx (m)

(5.28)

k=0

where ryx (m) is the crosscorrelation sequence of the input {x(n)} to the system with the output {y(n)} of the system, and rxx (m) is the autocorrelation sequence of the input signal. In the frequency domain, the corresponding relationship is Syx (ω) = H (ω)Sxx(ω) = H (ω)|X(ω)|2 Hence H (ω) =

Syx (ω) Syx (ω) = Sxx (ω) |X(ω)|2

(5.29)

These relations suggest that the impulse response {h(n)} or the frequency response of an unknown system can be determined (measured) by crosscorrelating the input sequence {x(n)} with the output sequence {y(n)}, and then solving the deconvolution problem in (5.28) by means of the recursive equation in (5.27). Alternatively, we could simply compute the Fourier transform of (5.28) and determine the frequency response given by (5.29). Furthermore, if we select the input sequence {x(n)} such that its autocorrelation sequence {rxx (n)}, is a unit sample sequence, or equivalently, that its spectrum is flat (constant) over the passband of H (ω), the values of the impulse response {h(n)} are simply equal to the values of the crosscorrelation sequence {ryx (n)}. In general, the crosscorrelation method described above is an effective and practical method for system identification.

5.4

Homomorphic Deconvolution

The complex cepstrum is a useful tool for performing deconvolution in some applications such as seismic signal processing. To describe this method, let us

368

Frequency-Domain Analysis of LTI Systems

{y(n)}

Y(z) z-Transform

Complex logarithm

lnY(z) Cy(z)

Inverse z-transform

{cy(z)}

Figure 5.4 Homomorphic system for obtaining the cepstrum {cy (n)}

of the sequence {y(n)}.

us suppose that {y(n)} is the output sequence of a linear time-invariant system which is excited by the input sequence {x(n)}. Then Y (z) = X(z)H (z)

(5.30)

where H (z) is the system function. The logarithm of Y (z) is Cy (z) = ln Y (z) = ln X(z) + ln H (z)

(5.31)

= Cx (z) + Ch (z) Consequently, the complex cepstrum of the output sequence {y(n)} is expressed as the sum of the cepstrum of {x(n)} and {h(n)}, that is, cy (n) = cx (n) + ch (n)

(5.32)

Thus we observe that convolution of the two sequences in the time domain corresponds to the summation of the cepstrum sequences in the cepstral domain. The system for performing these transformations is called a homormorphic systemand is illustrated in Fig. 5.4. In some applications, such as seismic signal processing and speech signal processing, the characteristics of the cepstral sequences {cx (n)} and {ch (n)} are sufficiently different so that they can be separated in the cepstral domain. Specifically, suppose that {ch (n)} has its main components (main energy) in the vicinity of small values of n, whereas {cx (n)} has its components concentrated at large values of n. We may say that {ch (n)} is “lowpass” and {cx (n)} is “highpass.” We can then separate {ch (n)} from {cx (n)} using appropriate “lowpass” and “highpass” windows, as illustrated in Fig. 5.5. Thus cˆh(n)

cy(n) = cx(n) + ch(n)

wlp(n) cˆx(n)

cy(n) = cx(n) + ch(n)

Figure 5.5

Separating the two cepstral components by “lowpass” and “highpass” windows.

whp(n)

369

Frequency-Domain Analysis of LTI Systems

Cˆ x(x)

cˆx(n) z-Transform cˆh(n)

Cˆ h(x)

ˆ X(z) Complex exponential

ˆ H(z)

Inverse z-transform

x(n) ˆ ˆ h(n)

Inverse homomorphic system for recovering the sequences {x(n)} and {h(n)} from the corresponding cepstra.

Figure 5.6

cˆ h (n) = cy (n)wlp (n)

(5.33)

cˆ x (n) = cy (n)whp (n)

(5.34)

and where

wlp (n) =

whp (n) =

1, 0,

|n| ≤ N1 otherwise

(5.35)

0, 1,

|n| ≤ N1 |n| > N1

(5.36)

Once we have separated the cepstrum sequences {ˆch (n)} and {ˆcx (n)} by windowing, ˆ the sequences {x(n)} ˆ and {h(n)} are obtained by passing {ˆch (n)} and {ˆcx (n)} through the inverse homomorphic system, shown in Fig. 5.6. In practice, a digital computer would be used to compute the cepstrum of the sequence {y(n)}, to perform the windowing functions, and to implement the inverse homomorphic system shown in Fig. 5.6. In place of the z-transform and inverse ztransform, we would substitute a special form of the Fourier transform and its inverse. This special form is called the discrete Fourier transform.

6

Summary and References In this chapter we considered the frequency-domain characteristics of LTI systems. We showed that an LTI system is characterized in the frequency domain by its frequency response function H (ω), which is the Fourier transform of the impulse response of the system. We also observed that the frequency response function determines the effect of the system on any input signal. In fact, by transforming the input signal into the frequency domain, we observed that it is a simple matter to determine the effect of the system on the signal and to determine the system output. When viewed in the frequency domain, an LTI system performs spectral shaping or spectral filtering on the input signal. The design of some simple IIR filters was also considered in this chapter from the viewpoint of pole–zero placement. By means of this method, we were able to design simple digital resonators, notch filters, comb filters, all-pass filters, and digital sinusoidal generators. Digital sinusoidal generators find use in frequency synthesis applications. A comprehensive treatment of frequency synthesis techniques is given in the text edited by Gorski-Popiel (1975).

370

Frequency-Domain Analysis of LTI Systems

Finally, we characterized LTI systems as either minimum-phase, maximum-phase, or mixed-phase, depending on the position of their poles and zeros in the frequency domain. Using these basic characteristics of LTI systems, we considered practical problems in inverse filtering, deconvolution, and system identification. We concluded with the description of a deconvolution method based on cepstral analysis of the output signal from a linear system. A vast amount of technical literature exists on the topics of inverse filtering, deconvolution, and system identification. In the context of communications, system identification and inverse filtering as they relate to channel equalization are treated in the book by Proakis (2001). Deconvolution techniques are widely used in seismic signal processing. For reference, we suggest the papers by Wood and Treitel (1975), Peacock and Treitel (1969), and the books by Robinson and Treitel (1978, 1980). Homomorphic deconvolution and its applications to speech processing aretreated in the book by Oppenheim and Schafer (1989).

Problems 1 The following input–output pairs have been observed during the operation of various systems: T1

(a) x(n) = ( 21 )n −→ y(n) = ( 18 )n T2

(b) x(n) = ( 21 )n u(n) −→ y(n) = ( 18 )n u(n) T3

(c) x(n) = ej π/5 −→ y(n) = 3ej π/5 T4

(d) x(n) = ej π/5 u(n) −→ y(n) = 3ej π/5u(n) T5

(e) x(n) = x(n + N1 ) −→ y(n) = y(n + N2 ), 2

N1 = N2 ,

N1 , N2 prime

Determine their frequency response if each of the above systems is LTI. (a) Determine and sketch the Fourier transform WR (ω) of the rectangular sequence

1, 0≤n≤M wR (n) = 0, otherwise (b) Consider the triangular sequence  n, wT (n) = M − n, 0,

0 ≤ n ≤ M/2 M/2 < 2 ≤ M otherwise

Determine and sketch the Fourier transform WT (ω) of wT (n) by expressing it as the convolution of a rectangular sequence with itself. (c) Consider the sequence   1 2π n 1 + cos wR (n) wc (n) = 2 M Determine and sketch Wc (ω) by using WR (ω).

371

Frequency-Domain Analysis of LTI Systems

3 Consider an LTI system with impulse response h(n) = ( 21 )n u(n). (a) Determine and sketch the magnitude and phase response |H (ω)| and H (ω), respectively. (b) Determine and sketch the magnitude and phase spectra for the input and output signals for the following inputs: 3πn , −∞ < n < ∞ 10 2. x(n) = {. . . , 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0, 1, . . .} 1. x(n) = cos



4

Determine and sketch the magnitude and phase response of the following systems: (a) y(n) = 21 [x(n) + x(n − 1)] (b) y(n) = 21 [x(n) − x(n − 1)] (c) y(n) = 21 [x(n + 1) − x(n − 1)] (d) y(n) = 21 [x(n + 1) + x(n − 1)] (e) y(n) = 21 [x(n) + x(n − 2)] (f) y(n) = 21 [x(n) − x(n − 2)] (g) y(n) = 13 [x(n) + x(n − 1) + x(n − 2)] (h) y(n) = x(n) − x(n − 8) (i) y(n) = 2x(n − 1) − x(n − 2) (j) y(n) = 41 [x(n) + x(n − 1) + x(n − 2) + x(n − 3)] (k) y(n) = 18 [x(n) + 3x(n − 1) + 3x(n − 2) + x(n − 3)] (l) y(n) = x(n − 4) (m) y(n) = x(n + 4) (n) y(n) = 41 [x(n) − 2x(n − 1) + x(n − 2)]

5 An FIR filter is described by the difference equation y(n) = x(n) + x(n − 10) (a) Compute and sketch its magnitude and phase response. (b) Determine its response to the inputs π π π 1. x(n) = cos n + 3 sin n+ , −∞ < n < ∞ 10 3 10   2π π 2. x(n) = 10 + 5 cos n+ , −∞ < n < ∞ 5 2

372

Frequency-Domain Analysis of LTI Systems

6

Determine the transient and steady-state responses of the FIR filter shown in Fig. P6 to the input signal x(n) = 10ej πn/2 u(n). Let b = 2 and y(−1) = y(−2) = y(−3) = y(−4) = 0. z −1

x(n)

z −1

z −1

z −1

b

− +

Figure P6 y(n)

7 Consider the FIR filter y(n) = x(n) + x(n − 4) (a) Compute and sketch its magnitude and phase response. (b) Compute its response to the input x(n) = cos

π π n + cos n, 2 4

−∞ < n < ∞

(c) Explain the results obtained in part (b) in terms of the magnitude and phase responses obtained in part (a). 8

Determine the steady-state and transient responses of the system 1 y(n) = [x(n) − x(n − 2)] 2 to the input signal x(n) = 5 + 3 cos

π 2

 n + 60◦ ,

−∞ < n < ∞

9 From our discussions it is apparent that an LTI system cannot produce frequencies at its output that are different from those applied in its input. Thus, if a system creates “new” frequencies, it must be nonlinear and/or time varying. Determine the frequency content of the outputs of the following systems to the input signal x(n) = A cos

π n 4

(a) y(n) = x(2n) (b) y(n) = x 2 (n) (c) y(n) = (cos π n)x(n)

373

Frequency-Domain Analysis of LTI Systems

10 Determine and sketch the magnitude and phase response of the systems shown in Fig. P10(a) through (c).

x(n)

z −1

+

+

1 2

y(n)

+ (a)

x(n)

z −1

+

+

1 2

y(n)

− (b)

x(n)

z −1

+

z −1

+

z −1

+

1 8

y(n)

(c)

Figure P10

11 Determine the magnitude and phase response of the multipath channel y(n) = x(n) + x(n − M)

12

At what frequencies does H (ω) = 0? Consider the filter y(n) = 0.9y(n − 1) + bx(n)

(a) Determine b so that |H (0)| = 1. √ (b) Determine the frequency at which |H (ω)| = 1/ 2. (c) Is this filter lowpass, bandpass, or highpass? (d) Repeat parts (b) and (c) for the filter y(n) = −0.9y(n − 1) + 0.1x(n). 13 Harmonic distortion in digital sinusoidal generators An ideal sinusoidal generator produces the signal x(n) = cos 2πf0 n,

−∞ < n < ∞

which is periodic with fundamental period N if f0 = k0 /N and k0 , N are relatively prime numbers. The spectrum of such a “pure” sinusoid consist of two lines at k = k0 and k = N − k0 (we limit ourselves in the fundamental interval 0 ≤ k ≤ N − 1). In practice, the approximations made in computing the samples of a sinusoid of relative frequency f0 result in a certain amount of power falling into other frequencies. This spurious power results in distortion, which is referred to as harmonic distortion.

374

Frequency-Domain Analysis of LTI Systems

Harmonic distortion is usually measured in terms of the total harmonic distortion (THD), which is defined as the ratio THD =

spurious harmonic power total power

(a) Show that THD = 1 − 2

|ck0 |2 Px

where ck0 =

N−1 1  x(n)e−j (2π/N)k0 n N n=0

N−1 1  Px = |x(n)|2 N n=0

(b) By using the Taylor approximation cos φ = 1 −

φ2 φ4 φ6 + − + ··· 2! 4! 6!

compute one period of x(n) for f0 = 1/96, 1/32, 1/256 by increasing the number of terms in the Taylor expansion from 2 to 8. (c) Compute the THD and plot the power density spectrum for each sinusoid in part (b) as well as for the sinusoids obtained using the computer cosine function. Comment on the results. 14 Measurement of the total harmonic distortion in quantized sinusoids Let x(n) be a periodic sinusoidal signal with frequency f0 = k/N , that is, x(n) = sin 2πf0 n (a) Write a computer program that quantizes the signal x(n) into b bits or equivalently into L = 2b levels by using rounding. The resulting signal is denoted by xq (n). (b) For f0 = 1/50 compute the THD of the quantized signals xq (n) obtained by using b = 4, 6, 8, and 16 bits. (c) Repeat part (b) for f0 = 1/100. (d) Comment on the results obtained in parts (b) and (c). 15 Consider the discrete-time system y(n) = ay(n − 1) + (1 − a)x(n),

n≥0

where a = 0.9 and y(−1) = 0.

375

Frequency-Domain Analysis of LTI Systems

(a) Compute and sketch the output yi (n) of the system to the input signals xi (n) = sin 2πfi n, where f1 = 41 , f2 = 15 , f3 =

1 10 ,

f4 =

0 ≤ n ≤ 100

1 20 .

(b) Compute and sketch the magnitude and phase response of the system and use these results to explain the response of the system to the signals given in part (a). 16

Consider an LTI system with impulse response h(n) = ( 31 )|n| . (a) Determine and sketch the magnitude and phase response H (ω)| and ⭿H (ω), respectively. (b) Determine and sketch the magnitude and phase spectra for the input and output signals for the following inputs: 3πn 1. x(n) = cos , −∞ < n < ∞ 8 2. x(n) = {. . . , −1, 1, −1, 1, −1, 1, −1, 1, −1, 1, −1, 1, . . .} ↑

17 Consider the digital filter shown in Fig. P17. (a) Determine the input–output relation and the impulse response h(n). (b) Determine and sketch the magnitude |H (ω)| and the phase response ⭿H (ω) of the filter and find which frequencies are completely blocked by the filter. (c) When ω0 = π/2, determine the output y(n) to the input x(n) = 3 cos

π 3

 n + 30◦ ,

x(n)

−∞ < n < ∞

z −1

Figure P17

z −1

+

y(n)

a = −2 cos ω0

18 Consider the FIR filter y(n) = x(n) − x(n − 4) (a) Compute and sketch its magnitude and phase response. (b) Compute its response to the input x(n) = cos

π π n + cos n, 2 4

−∞ < n < ∞

(c) Explain the results obtained in part (b) in terms of the answer given in part (a).

376

Frequency-Domain Analysis of LTI Systems

19

Determine the steady-state response of the system y(n) =

1 [x(n) − x(n − 2)] 2

to the input signal x(n) = 5 + 3 cos 20

π 2

 n + 60◦ + 4 sin(π n + 45◦ ),

−∞ < n < ∞

Recall from Problem 9 that an LTI system cannot produce frequencies at its output that are different from those applied in its input. Thus if a system creates “new” frequencies, it must be nonlinear and/or time varying. Indicate whether the following systems are nonlinear and/or time varying and determine the output spectra when the input spectrum is

1, |ω| ≤ π/4 X(ω) = 0, π/4 ≤ |ω| ≤ π (a) y(n) = x(2n) (b) y(n) = x 2 (n) (c) y(n) = (cos π n)x(n)

21 Consider an LTI system with impulse response h(n) =

 n  π  1 cos n u(n) 4 4

(a) Determine its system function H (z). (b) Is it possible to implement this system using a finite number of adders, multipliers, and unit delays? If yes, how? (c) Provide a rough sketch of |H (ω)| using the pole–zero plot. (d) Determine the response of the system to the input 1 x(n) = ( )n u(n) 4 22 An FIR filter is described by the difference equation y(n) = x(n) − x(n − 6) (a) Compute and sketch its magnitude and phase response. (b) Determine its response to the inputs π π π 1. x(n) = cos n + 3 sin n+ , −∞ < n < ∞ 10  3  10 2π π 2. x(n) = 5 + 6 cos n+ , −∞ < n < ∞ 5 2

377

Frequency-Domain Analysis of LTI Systems

23 The frequency response of an ideal bandpass filter is given by   0,     H (ω) = 1,      0,

|ω| ≤

π 8

π 3π < |ω| < 8 8 3π ≤ |ω| ≤ π 8

(a) Determine its impulse response (b) Show that this impulse response can be expressed as the product of cos(nπ/4) and the impulse response of a lowpass filter. 24 Consider the system described by the difference equation y(n) =

1 1 y(n − 1) + x(n) + x(n − 1) 2 2

(a) Determine its impulse response. (b) Determine its frequency response: 1. From the impulse response 2. From the difference equation (c) Determine its response to the input x(n) = cos

π 2

n+

π , 4

−∞ < n < ∞

25 Sketch roughly the magnitude |H (ω)| of the Fourier transforms corresponding to the pole–zero patterns of systems given in Fig. P25. Unit circle

Double pole

Pole at 0.9e jθ

Unit circle

Double zero (a)

0.9

(b)

0.9

1

Figure P25

378

8th-order pole

(c)

(d)

Frequency-Domain Analysis of LTI Systems

26

Design an FIR filter that completely blocks the frequency ω0 = π/4 and then compute its output if the input is  π  x(n) = sin n u(n) 4

27

for n = 0, 1, 2, 3, 4. Does the filter fulfill your expectations? Explain. A digital filter is characterized by the following properties: 1. It is highpass and has one pole and one zero. 2. The pole is at a distance r = 0.9 from the origin of the z-plane. 3. Constant signals do not pass through the system. (a) Plot the pole–zero pattern of the filter and determine its system function H (z). (b) Compute the magnitude response |H (ω)| and the phase response ⭿H (ω) of the filter. (c) Normalize the frequency response H (ω) so that |H (π )| = 1. (d) Determine the input–output relation (difference equation) of the filter in the time domain. (e) Compute the output of the system if the input is x(n) = 2 cos

π 6

 n + 45◦ ,

−∞ < n < ∞

(You can use either algebraic or geometrical arguments.) 28 A causal first-order digital filter is described by the system function H (z) = b0

1 + bz−1 1 + az−1

(a) Sketch the direct form I and direct form II realizations of this filter and find the corresponding difference equations. (b) For a = 0.5 and b = −0.6, sketch the pole–zero pattern. Is the system stable? Why? (c) For a = −0.5 and b = 0.5, determine b0 , so that the maximum value of |H (ω)| is equal to 1. (d) Sketch the magnitude response |H (ω)| and the phase response ⭿H (ω) of the filter obtained in part (c). (e) In a specific application it is known that a = 0.8. Does the resulting filter amplify high frequencies or low frequencies in the input? Choose the value of b so as to improve the characteristics of this filter (i.e., make it a better lowpass or a better highpass filter). 29 Derive the expression for the resonant frequency of a two-pole filter with poles at p1 = rej θ and p2 = p1∗ , given by (4.25).

379

Frequency-Domain Analysis of LTI Systems

30 Determine and sketch the magnitude and phase responses of the Hanning filter characterized by the (moving average) difference equation y(n) =

1 1 1 x(n) + x(n − 1) + x(n − 2) 4 2 4

31 A causal LTI system excited by the input 1 x(n) = ( )n u(n) + u(−n − 1) 4 produces an output y(n) with z-transform Y (z) =

− 43 z−1 (1 − 41 z−1 )(1 + z−1 )

(a) Determine the system function H (z) and its ROC. (b) Determine the output y(n) of the system. (Hint: Pole cancellation increases the original ROC.) 32 Determine the coefficients of a linear-phase FIR filter y(n) = b0 x(n) + b1 x(n − 1) + b2 x(n − 2) such that: (a) It rejects completely a frequency component at ω0 = 2π/3. (b) Its frequency response is normalized so that H (0) = 1. (c) Compute and sketch the magnitude and phase response of the filter to check if it satisfies the requirements. 33 Determine the frequency response H (ω) of the following moving average filters. (a) y(n) =

M  1 x(n − k) 2M + 1 k=−M

1 1 x(n + M) + (b) y(n) = 4M 2M 34

M−1 

x(n − k) +

k=−M+1

1 x(n − M) 4M

Which filter provides better smoothing? Why? Compute the magnitude and phase response of a filter with system function H (z) = 1 + z−1 + z−2 + · · · + z−8

If the sampling frequency is Fs = 1 kHz, determine the frequencies of the analog sinusoids that cannot pass through the filter. 35 A second-order system has a double pole at p1,2 = 0.5 and two zeros at z1,2 = e±j 3π/4 Using geometric arguments, choose the gain G of the filter so that |H (0)| = 1.

380

Frequency-Domain Analysis of LTI Systems

36

In this problem we consider the effect of a single zero on the frequency response of a system. Let z = rej θ be a zero inside the unit circle (r < 1). Then Hz (ω) = 1 − rej θ e−j ω = 1 − r cos(ω − θ) + j r sin(ω − θ) (a) Show that the magnitude response is |Hz (ω)| = [1 − 2r cos(ω − θ) + r 2 ]1/2 or, equivalently, 20 log10 |Hz (ω)| = 10 log10 [1 − 2r cos(ω − θ) + r 2 ] (b) Show that the phase response is given as z (ω) = tan−1

r sin(ω − θ) 1 − r cos(ω − θ)

(c) Show that the group delay is given as τg (ω) =

r 2 − r cos(ω − θ) 1 + r 2 − 2r cos(ω − θ)

(d) Plot the magnitude |H (ω)|dB , the phase (ω) and the group delay τg (ω) for r = 0.7 and θ = 0, π/2, and π . 37

In this problem we consider the effect of a single pole on the frequency response of a system. Hence, we let Hp (ω) =

1 , 1 − rej θ e−j ω

r 2B , we have X(F ) = Xa (F )/T for |F | ≤ Fs /2. Therefore, the output of the system in Figure 2.1 is given by  H (F )Xa (F ), |F | ≤ Fs /2 Ya (F ) = H (F )X(F )Ga (F ) = (2.8) 0, |F | > Fs /2 To assure that ya (t) = yˆ a (t), we should choose the discrete-time system so that  Ha (F ), |F | ≤ Fs /2 H (F ) = (2.9) 0, |F | > Fs /2

409

Sampling and Reconstruction of Signals

We note that, in this special case, the cascade connection of the A/D converter (linear time-varying system), an LTI (linear time-invariant) system, and the D/A converter (linear time-varying system) is equivalent to a continuous-timeLTI system. This important result provides the theoretical basis for the discrete-time filtering of continuous-time signals. These concepts are illustrated in the following examples. EXAMPLE 2.1

Simulation of an analog integrator

Consider the analog integrator circuit shown in Figure 2.4(a). given by RC

Its input–output relation is

dya (t) + ya (t) = xa (t) dt

Taking the Fourier transform of both sides, we can show that the frequency response of the integrator is Ha (F ) =

Ya (F ) 1 = , Xa (F ) 1 + j F /Fc

Fc =

1 2πRC

Evaluating the inverse Fourier transform yields the impulse response

ha (t) = Ae−At u(t),

A=

1 RC

Clearly the impulse response ha (t) is a nonbandlimited signal. We now define a discrete-time system by sampling the continuous-time impulse response as follows: h(n) = ha (nT ) = A(e−AT )n u(n)

We say that the discrete-time system is obtained from the continuous-time system through an impulse-invariance transformation. The system function and the difference equation of the discrete-time system are

H (z) =

∞  n=0

A(e−AT )n z−n =

1 1 − e−AT z−1

y(n) = e−AT y(n − 1) + Ax(n) The system is causal and has a pole p = e−AT . Since A > 0, |p| < 1 and the system is always stable. The frequency response of the system is obtained by evaluating H (z) for z = ej 2π F /Fs . Figure 2.4(b) shows the magnitude frequency responses of the analog integrator and the discrete-time simulator for Fs =50, 100, 200, and 1000 Hz. We note that the effects of aliasing, caused by the sampling of ha (t), become negligible only for sampling frequencies larger than 1 kHz. The discrete-time implementation is accurate for input signals with bandwidth much less than the sampling frequency.

410

Sampling and Reconstruction of Signals

R

xa(t )

y(n)

x(n) ha(t )

ya(t )

C

A _ z 1

h(n) t

T

0

eAT

Fs = 50 Hz

1 0.8 Magnitude

0.6 Fs = 100 Hz

0.4 Ha(F )

0.2

Fs = 200 Hz

Fs = 1 KHz 0

20

40

60

80 100 120 140 160 180 200 Frequency F (Hz)

Discrete-time implementation of an analog integrator using impulse response sampling. The approximation is satisfactory when the bandwidth of the input signal is much less than the sampling frequency.

Figure 2.4

EXAMPLE 2.2

Ideal bandlimited differentiator

The ideal continuous-time differentiator is defined by

ya (t) =

dxa (t) dt

(2.10)

and has frequency response function

Ha (F ) =

Ya (F ) = j 2πF Xa (F )

(2.11)

For processing bandlimited signals, it is sufficient to use the ideal bandlimited differentiator defined by  j 2πF, |F | ≤ Fc Ha (F ) = (2.12) 0, |F | > Fc If we choose Fs = 2Fc , we can define an ideal discrete time differentiator by H (F ) = Ha (F ) = j 2πF,

|F | ≤ Fs /2

(2.13)

411

Sampling and Reconstruction of Signals

 Since by definition H (F ) = k Ha (F −kFs ), we have h(n) = ha (nT ). In terms of ω = 2πF /Fs , H (ω) is periodic with period 2π . Therefore, the discrete-time impulse response is given by  π πn cos πn − sin πn 1 H (ω)ej ωn dω = (2.14) h(n) = 2π −π πn2 T or in a more compact form

 h(n) =

0, cos π n , nT

n=0 n = 0

(2.15)

The magnitude and phase responses of the continuous-time and discrete-time ideal differentiators are shown in Figure 2.5.

|Ha(F)|





Fs 2

Fs 2

π 2

Fs 2

F

Fs 2

F

⭿Ha(F)

π − 2 (a) |H(ω)|

−2π

−π

π

π 2 −2π



ω

⭿H(ω)

−π

π − 2

π



ω

(b) Frequency responses of the ideal bandlimited continuoustime differentiator (a) and its discrete-time counterpart (b).

Figure 2.5

412

Sampling and Reconstruction of Signals

EXAMPLE 2.3

Fractional delay

A continuous-time delay system is defined by ya (t) = xa (t − td )

(2.16)

for any td > 0. Although the concept is simple, its practical implementation is quite complicated. If xa (t) is bandlimited and sampled at the Nyquist rate, we obtain y(n) = ya (nT ) = xa (nT − td ) = xa [(n − )T ] = x(n − )

(2.17)

where  = td /T . If  is an integer, delaying the sequence x(n) is a simple process. For noninteger values of , the delayed value of x(n) would lie somewhere between two samples. However, this value is unavailable and the only way to generate an appropriate value is by ideal bandlimited interpolation. One way to approach this problem is by considering the frequency response (2.18) Hid (ω) = e−j ω of the delay system in (2.17) and its impulse response  π 1 sin π(n − ) hid (n) = H (ω)ej ωn dω = 2π −π π(n − )

(2.19)

When the delay  assumes integer values, hid (n) reduces to δ(n − ) because the sin function is sampled at the zero crossings. When  is noninteger, hid (n) is infinitely long because the sampling times fall between the zero crossings. Unfortunately, the ideal impulse response for fractional delay systems is noncausal and has infinite duration. Therefore, the frequency response (2.18) has to be approximated with realizable FIR or IIR filters. More details about fractional delay filter design can be found in Laakso et al. (1996).

3

Analog-to-Digital and Digital-to-Analog Converters In the previous section we assumed that the A/D and D/A converters in the processing of continuous-time signals are ideal. The one implicit assumption that we have made in the discussion on the equivalence of continuous-time and discrete-time signal processing is that the quantization error in analog-to-digital conversion and round-off errors in digital signal processing are negligible. These issues are further discussed in this section. However, we should emphasize that analog signal processing operations cannot be done very precisely either, since electronic components in analog systems have tolerances and they introduce noise during their operation. In general, a digital system designer has better control of tolerances in a digital signal processing system than an analog system designer who is designing an equivalent analog system. The discussion in Section 1 focused on the conversion of continuous-time signals to discrete-time signals using an ideal sampler and ideal interpolation. In this section we deal with the devices for performing these conversions from analog to digital.

3.1

Analog-to-Digital Converters

Recall that the process of converting a continuous-time (analog) signal to a digital sequence that can be processed by a digital system requires that we quantize the sampled values to a finite number of levels and represent each level by a number of bits.

413

Sampling and Reconstruction of Signals

Convert command

S/H control

A/D converter

Samplehold Analog preamp

To computer or communication channel

Buffer or bus

Status (a) Tracking in "sample"

Input

H

S

Holding

H

S

H

S

H

S

H

S

S H

S/H output (b)

(a) Block diagram of basic elements of an A/D converter; (b) time-domain response of an ideal S/H circuit.

Figure 3.1

Figure 3.1(a) shows a block diagram of the basic elements of an A/D converter . In this section we consider the performance requirements for these elements. Although we focus mainly on ideal system characteristics, we shall also mention some key imperfections encountered in practical devices and indicate how they affect the performance of the converter. We concentrate on those aspects that are more relevant to signal processing applications. The practical aspects of A/D converters and related circuitry can be found in the manufacturers’ specifications and data sheets. In practice, the sampling of an analog signal is performed by a sample-and-hold (S/H) circuit. The sampled signal is then quantized and converted to digital form. Usually, the S/H is integrated into the A/D converter. The S/H is a digitally controlled analog circuit that tracks the analog input signal during the sample mode, and then holds it fixed during the hold mode to the instantaneous value of the signal at the time the system is switched from the sample mode to the hold mode. Figure 3.1(b) shows the time-domain response of an ideal S/H circuit (i.e., an S/H that responds instantaneously and accurately). The goal of the S/H is to continuously sample the input signal and then to hold that value constant as long as it takes for the A/D converter to obtain its digital representation. The use of an S/H allows the A/D converter to operate more slowly compared to the time actually used to acquire the sample. In the absence of an S/H, the input signal must not change by more than one-half of the quantization step during the conversion, which may be an impractical constraint. Consequently, the S/H is crucial in high-resolution (12 bits per sample or higher) digital conversion of signals that have large bandwidths (i.e., they change very rapidly).

414

Sampling and Reconstruction of Signals

An ideal S/H introduces no distortion in the conversion process and is accurately modeled as an ideal sampler. However, time-related degradations such as errors in the periodicity of the sampling process (“jitter”), nonlinear variations in the duration of the sampling aperture, and changes in the voltage held during conversion (“droop”) do occur in practical devices. The A/D converter begins the conversion after it receives a convert command. The time required to complete the conversion should be less than the duration of the hold mode of the S/H. Furthermore, the sampling period T should be larger than the duration of the sample mode and the hold mode. In the following sections we assume that the S/H introduces negligible errors and we focus on the digital conversion of the analog samples.

3.2

Quantization and Coding

The basic task of the A/D converter is to convert a continuous range of input amplitudes into a discrete set of digital code words. This conversion involves the processes of quantization and coding. Quantization is a nonlinear and noninvertible process that maps a given amplitude x(n) ≡ x(nT ) at time t = nT into an amplitude xk , taken from a finite set of values. The procedure is illustrated in Fig. 3.2(a), where the signal amplitude range is divided into L intervals Ik = {xk < x(n) ≤ xk+1 },

k = 1, 2, . . . , L

(3.1)

by the L + 1 decision levels x1 , x2 , . . . , xL+1 . The possible outputs of the quantizer (i.e., the quantization levels) are denoted as xˆ 1 , xˆ 2 , . . . , xˆ L . The operation of the quantizer is defined by the relation xq (n) ≡ Q[x(n)] = xˆ k , Quantization levels

Decision levels

if x(n) ∈ Ik

(3.2)

Ik



… x^3

x3

x4

x^4



x5

xk

x^k xk + 1

Instantaneous amplitude (a) x1 = −∞ x^1 x2 −4∆

x^2 x3 −3∆

x^3 −2∆

x4 x^4

x5

−∆

x^5

x6

0

x^6 ∆

x7

x^7 2∆

x8

x^8

x9 = ∞

3∆

Instantaneous amplitude Range of quantizer (b)

Figure 3.2

Quantization process and an example of a midtread

quantizer.

415

Sampling and Reconstruction of Signals

In most digital signal processing operations the mapping in (3.2) is independent of n (i.e., the quantization is memoryless and is simply denoted as xq = Q[x]). Furthermore, in signal processing we often use uniform or linear quantizers defined by xˆ k+1 − xˆ k = ,

k = 1, 2, . . . , L − 1

xk+1 − xk = ,

for finite xk , xk+1

(3.3)

where  is the quantizer step size. Uniform quantization is usually a requirement if the resulting digital signal is to be processed by a digital system. However, in transmission and storage applications of signals such as speech, nonlinear and timevariant quantizers are frequently used. If zero is assigned a quantization level, the quantizer is of the midtread type. If zero is assigned a decision level, the quantizer is called a midrise type. Figure 3.2(b) illustrates a midtread quantizer with L = 8 levels. In theory, the extreme decision levels are taken as x1 = −∞ and xL+1 = ∞, to cover the total dynamic range of the input signal. However, practical A/D converters can handle only a finite range. Hence we define the range R of the quantizer by assuming that I1 = IL = . For example, the range of the quantizer shown in Fig. 3.2(b) is equal to 8. In practice, the term full-scale range (FSR) is used to describe the range of an A/D converter for bipolar signals (i.e., signals with both positive and negative amplitudes). The term full scale (FS) is used for unipolar signals. It can be easily seen that the quantization error eq (n) is always in the range −/2 to /2:   (3.4) − < eq (n) ≤ 2 2 In other words, the instantaneous quantization error cannot exceed half of the quantization step. If the dynamic range of the signal, defined as xmax − xmin , is larger than the range of the quantizer, the samples that exceed the quantizer range are clipped, resulting in a large (greater than /2) quantization error. The operation of the quantizer is better described by the quantization characteristic function, illustrated in Fig. 3.3 for a midtread quantizer with eight quantization levels. This characteristic is preferred in practice over the midriser because it provides an output that is insensitive to infinitesimal changes of the input signal about zero. Note that the input amplitudes of a midtread quantizer are rounded to the nearest quantization levels. The coding process in an A/D converter assigns a unique binary number to each quantization level. If we have L levels, we need at least L different binary numbers. With a word length of b + 1 bits we can represent 2b+1 distinct binary numbers. Hence we should have 2b+1 ≥ L or, equivalently, b + 1 ≥ log2 L. Then the step size or the resolution of the A/D converter is given by = where R is the range of the quantizer.

416

R 2b+1

(3.5)

Sampling and Reconstruction of Signals

Output x^ = Q[x] Quantization levels

3∆ 2∆

−∆ 2 −

9∆ 2



7∆ 2



5∆ 2





3∆ 2

∆ 2 −∆

−2∆

3∆ 2

Two's-complement code words 011 Decision levels 010 001 000 x 5∆ 7∆ 9∆ 111 Input 110 2 2 2 101 100

−3∆ −4∆

Range R = FSR (Peak-to-peak range)

Figure 3.3

Example of a midtread quantizer.

There are various binary coding schemes, each with its advantages and disadvantages. Table 1 illustrates some existing schemes for 3-bit binary coding. The two’s-complement representation is used in most digital signal processors. Thus it is convenient to use the same system to represent digital signals because we can operate on them directly without any extra format conversion. In general, a (b + 1)-bit binary fraction of the form β0 β1 β2 · · · βb has the value −β0 · 20 + β1 · 2−1 + β2 · 2−2 + · · · + βb · 2−b if we use the two’s-complement representation. Note that β0 is the most significant bit (MSB) and βb is the least significant bit (LSB). Although the binary code used to represent the quantization levels is important for the design of the A/D converter and the subsequent numerical computations, it does not have any effect in the performance of the quantization process. Thus in our subsequent discussions we ignore the process of coding when we analyze the performance of A/D converters. The only degradation introduced by an ideal converter is the quantization error, which can be reduced by increasing the number of bits. This error, which dominates the performance of practical A/D converters, is analyzed in the next section. Practical A/D converters differ from ideal converters in several ways. Various degradations are usually encountered in practice. Specifically, practical A/D converters may have an offset error (the first transition may not occur at exactly + 21 LSB),

417

Sampling and Reconstruction of Signals

TABLE 1

Commonly Used Bipolar Codes Decimal Fraction

Positive Negative Sign + Two’s Offset One’s Number Reference Reference Magnitude Complement Binary Complement +7

+ 78

− 78

0111

0111

1111

0111

+6

− 68 − 58 − 48 − 38 − 28 − 18

0110

0110

1110

0110

0101

0101

1101

0101

0100

0100

1100

0100

0011

0011

1011

0011

0010

0010

1010

0010

+1

+ 68 + 58 + 48 + 38 + 28 + 18

0001

0001

1001

0001

0

0+

0−

0000

0000

1000

0000

0

0−

0+

1000

(0 0 0 0)

(1 0 0 0)

1111

−1

− 18

+ 18

1001

1111

0111

1110

−2

− 28

+ 28

1010

1110

0110

1101

−3

− 38 − 48 − 58 − 68 − 78 − 88

+ 38 + 48 + 58 + 68 + 78 + 88

1011

1101

0101

1100

1100

1100

0100

1011

1101

1011

0011

1010

1110

1010

0010

1001

1001

0001

1000

(1 0 0 0)

(0 0 0 0)

+5 +4 +3 +2

−4 −5 −6 −7 −8

1111

scale-factor (or gain) error (the difference between the values at which the first transition and the last transition occur is not equal to FS − 2LSB), and a linearity error (the differences between transition values are not all equal or uniformly changing). If the differential linearity error is large enough, it is possible for one or more code words to be missed. Performance data on commercially available A/D converters are specified in manufacturers’ data sheets.

3.3

Analysis of Quantization Errors

To determine the effects of quantization on the performance of an A/D converter, we adopt a statistical approach. The dependence of the quantization error on the characteristics of the input signal and the nonlinear nature of the quantizer make a deterministic analysis intractable, except in very simple cases. In the statistical approach, we assume that the quantization error is random in nature. We model this error as noise that is added to the original (unquantized) signal. If the input analog signal is within the range of the quantizer, the quantization error

418

Sampling and Reconstruction of Signals

Quantizer Q[x(n)]

x(n)

xq(n)

(a) Actual system

x(n)

Figure 3.4

+

xq(n) = x(n) + eq(n)

eq(n)

Mathematical model of quantization noise.

(b) Mathematical model

eq (n) is bounded in magnitude [i.e., |eq (n)| < /2], and the resulting error is called granular noise. When the input falls outside the range of the quantizer (clipping), eq (n) becomes unbounded and results in overload noise. This type of noise can result in severe signal distortion when it occurs. Our only remedy is to scale the input signal so that its dynamic range falls within the range of the quantizer. The following analysis is based on the assumption that there is no overload noise. The mathematical model for the quantization error eq (n) is shown in Fig. 3.4. To carry out the analysis, we make the following assumptions about the statistical properties of eq (n): 1. The error eq (n) is uniformly distributed over the range −/2 < eq (n) < /2. 2. The error sequence {eq (n)} is a stationary white noise sequence. In other words, the error eq (n) and the error eq (m) for m = n are uncorrelated. 3. The error sequence {eq (n)} is uncorrelated with the signal sequence x(n). 4. The signal sequence x(n) is zero mean and stationary. These assumptions do not hold, in general. However, they do hold when the quantization step size is small and the signal sequence x(n) traverses several quantization levels between two successive samples. Under these assumptions, the effect of the additive noise eq (n) on the desired signal can be quantified by evaluating the signal-to-quantization-noise (power) ratio (SQNR), which can be expressed on a logarithmic scale (in decibels or dB) as SQNR = 10 log10

Px Pn

(3.6)

where Px = σx2 = E[x 2 (n)] is the signal power and Pn = σe2 = E[eq2 (n)] is the power of the quantization noise. If the quantization error is uniformly distributed in the range (−/2, /2) as shown in Fig. 3.5, the mean value of the error is zero and the variance (the quantization noise power) is  Pn = σe2 =

/2

−/2

e2 p(e) de =

1 



/2

−/2

e2 de =

2 12

(3.7)

419

Sampling and Reconstruction of Signals

p(e)

1 ∆

Figure 3.5

− ∆ 2

Probability density function for the quantization error.

0

e

∆ 2

By combining (3.5) with (3.7) and substituting the result into (3.6), the expression for the SQNR becomes SQNR = 10 log

Px σx = 20 log Pn σe

(3.8)

R = 6.02b + 16.81 − 20 log dB σx

The last term in (3.8) depends on the range R of the A/D converter and the statistics of the input signal. For example, if we assume that x(n) is Gaussian distributed and the range of the quantizer extends from −3σx to 3σx (i.e., R = 6σx ), then less than 3 out of every 1000 input signal amplitudes would result in an overload on the average. For R = 6σx , (3.8) becomes SQNR

6.02b

1.25 dB

(3.9)

The formula in (3.8) is frequently used to specify the precision needed in an A/D converter. It simply means that each additional bit in the quantizer increases the signal-to-quantization-noise ratio by 6 dB. (It is interesting to note that the same result is derived for a sinusoidal signal using a deterministic approach.) However, we should bear in mind the conditions under which this result has been derived. Due to limitations in the fabrication of A/D converters, their performance falls short of the theoretical value given by (3.8). As a result, the effective number of bits may be somewhat less than the number of bits in the A/D converter. For instance, a 16-bit converter may have only an effective 14 bits of accuracy.

3.4

Digital-to-Analog Converters

In practice, D/A conversion is usually performed by combining a D/A converter with a sample-and-hold (S/H) followed by a lowpass (smoothing) filter, as shown in Fig. 3.6. The D/A converter accepts, at its input, electrical signals that correspond Digital input signal

Figure 3.6

420

Digitalto-analog converter

Sample and hold

Lowpass smoothing filter

Analog output signal

Basic operations in converting a digital signal into an analog signal.

Sampling and Reconstruction of Signals

Analog output voltage Ideal D/A

3∆ 2∆ ∆ 100 101 110 111 000 −∆

001 010 011 Input code words

−2∆ −3∆

Figure 3.7

−4∆

Ideal D/A converter characteristic.

to a binary word, and produces an output voltage or current that is proportional to the value of the binary word. Ideally, its input–output characteristic is as shown in Fig. 3.7 for a 3-bit bipolar signal. The line connecting the dots is a straight line through the origin. In practical D/A converters, the line connecting the dots may deviate from the ideal. Some of the typical deviations from ideal are offset errors, gain errors, and nonlinearities in the input–output characteristic. An important parameter of a D/A converter is its settling time, which is defined as the time required for the output of the D/A converter to reach and remain within a given fraction (usually, ± 21 LSB) of the final value, after application of the input code word. Often, the application of the input code word results in a high-amplitude transient, called a “glitch.” This is especially the case when two consecutive code words to the A/D differ by several bits. The usual way to remedy this problem is to use an S/H circuit designed to serve as a “deglitcher.” Hence the basic task of the S/H is to hold the output of the D/A converter constant at the previous output value until the new sample at the output of the D/A reaches steady state. Then it samples and holds the new value in the next sampling interval. Thus the S/H approximates the analog signal by a series of rectangular pulses whose height is equal to the corresponding value of the signal pulse. Figure 3.8 illustrates the response of an S/H to a discrete-time sinusoidal signal. As shown, the approximation, is basically a staircase function which takes the signal sample from the D/A converter and holds it for T seconds. When the next sample arrives, it jumps to the next value and holds it for T seconds, and so on. 1

Figure 3.8

Response of an S/H interpolator to a discrete-time sinusoidal signal.

S/H input

S/H output

0 −1 0

Analog signal 20

40

t

60

80

100

421

Sampling and Reconstruction of Signals

T |GBL(F)|

Figure 3.9

Frequency responses of sample-and-hold and the ideal bandlimited interpolator.

4 dB

|GSH(F)|



1 T



1 2T

0

1 2T

1 T

F

The interpolation function of the S/H system is a square pulse defined by  1, 0 ≤ t ≤ T (3.10) gSH (t) = 0, otherwise The frequency-domain characteristics are obtained by evaluating its Fourier transform  ∞ sin π F T −2πF (T /2) gSH (t)e−j 2πF t dt = T e GSH (F ) = (3.11) πF T −∞ The magnitude of GSH (F ) is shown in Figure 3.9, where we superimpose the magnitude response of the ideal bandlimited interpolator for comparison purposes. It is apparent that the S/H does not possess a sharp cutoff frequency characteristic. This is due to a large extent to the sharp transitions of its interpolation function gSH (t). As a consequence, the S/H passes undesirable aliased frequency components (frequencies above Fs /2) to its output. This effect is sometimes referred to as postaliasing. To remedy this problem, it is common practice to filter the output of the S/H by passing it through a lowpass filter, which highly attenuates frequency components above Fs /2. In effect, the lowpass filter following the S/H smooths its output by removing sharp discontinuities. Sometimes, the frequency response of the lowpass filter is defined by  π F T e2πF (T /2) , |F | ≤ F /2 s Ha (F ) = sin π F T (3.12) 0, |F | > Fs /2 to compensate for the sinx/x distortion of the S/H (aperture effect). The aperture effect attenuation compensation, which reaches a maximum of 2/π or 4 dB at F = Fs /2, is usually neglected. However, it may be introduced using a digital filter before the sequence is applied to the D/A converter. The half-sample delay introduced by the S/H cannot be compensated because we cannot design analog filters that can introduce a time advance.

4

Sampling and Reconstruction of Continuous-Time Bandpass Signals A continuous-time bandpass signal with bandwidth B and center frequency Fc has its frequency content in the two frequency bands defined by 0 < FL < |F | < FH , where Fc = (FL +FH )/2 (see Figure 4.1(a)). A naive application of the sampling theorem

422

Sampling and Reconstruction of Signals

would suggest a sampling rate Fs ≥ 2FH ; however, as we show in this section, there are sampling techniques that allow sampling rates consistent with the bandwidth B , rather than the highest frequency, FH , of the signal spectrum. Sampling of bandpass signals is of great interest in the areas of digital communications, radar, and sonar systems.

4.1

Uniform or First-Order Sampling

Uniform or first-order sampling is the typical periodic sampling introduced in Section 1. Sampling the bandpass signal in Figure 4.1(a) at a rate F s = 1/T produces a sequence x(n) = xa (nT ) with spectrum X(F ) =

∞ 1  Xa (F − kFs ) T

(4.1)

k=−∞

The positioning of the shifted replicas X(F − kFs ) is controlled by a single parameter, the sampling frequency Fs . Since bandpass signals have two spectral bands, in general, it is more complicated to control their positioning, in order to avoid aliasing, with the single parameter Fs . Integer Band Positioning. We initially restrict the higher frequency of the band to be an integer multiple of the bandwidth, that is, FH = mB (integer band positioning). The number m = FH /B , which is in general fractional, is known as the band position. Figures 4.1 (a) and 4.1 (d) show two bandpass signals with even (m = 4) and odd (m = 3) band positioning. It c an be easily seen from Figure 4.1(b) that, for integerpositioned bandpass signals, choosing Fs = 2B results in a sequence with a spectrum without aliasing. From Figure 4.1(c), we see that the original bandpass signal can be reconstructed using the reconstruction formula xa (t) =

∞ 

xa (nT )ga (t − nT )

(4.2)

n=−∞

where

sin π Bt (4.3) cos 2π Fc t π Bt is the inverse Fourier transform of the bandpass frequency gating function shown in Figure 4.1(c). We note that ga (t) is equal to the ideal interpolation function for lowpass signals [see (1.21)], modulated by a carrier with frequency Fc . It is worth noticing that, by properly choosing the center frequency Fc of Ga (F ), we can reconstruct a continuous-time bandpass signal with spectral bands centered at Fc = ±(kB + B/2), k = 0, 1, . . .. For k = 0 we obtain the equivalent baseband signal, a process known as down-conversion. A simple inspection of Figure 4.1 demonstrates that the baseband spectrum for m = 3 has the same spectral structure as the original spectrum; however, the baseband spectrum for m = 4 has been “inverted.” In general, when the band position is an even integer the baseband spectral images are inverted versions of the original ones. Distinguishing between these two cases is important in communications applications. ga (t) =

423

Sampling and Reconstruction of Signals

|Xa(F)| 1 -Fc

-2Fs

-Fs

-3B

-4B

FL Fc FH

0 (a)

-2B

-B

Nyquist zones

|X(F)|

1 _ T

1st

2nd

B 0 (b) Ga(F) T

-Fc

F

Fs

3rd

4th

3B

2B

0

2Fs

4B

F

3B Fc 4B

F

(c) |Xa(F)| 1 -Fc

FL Fc FH

0

F

(d) -2Fs

-4B

|X(F)|

-Fs

-3B

-2B

1 _ T -B

1st

0

2nd B

Fs

2B

2Fs

3rd 3B

4B

F

(e) Figure 4.1

Illustration of bandpass signal sampling for integer band positioning.

Arbitrary Band Positioning. Consider now a bandpass signal with arbitrarily positioned spectral bands, as shown in Figure 4.2. To avoid aliasing, the sampling frequency should be such that the (k − 1)th and kth shifted replicas of the “negative” spectral band do not overlap with the “positive” spectral band. From Figure 4.2(b) we see that this is possible if there is an integer k and a sampling frequency Fs that satisfy the following conditions:

424

2FH ≤ kFs

(4.4)

(k − 1)Fs ≤ 2FL

(4.5)

Sampling and Reconstruction of Signals

|Xa(F)| B

B 1

−Fc

FL Fc FH

0

F

(a) |X(F)|

−Fc

kth replica

(k−1)th replica

1 _ T 0 (k−1)Fs

Fc

F

2FL 2FH kFs (b) Figure 4.2

Illustration of bandpass signal sampling for arbitrary band positioning.

which is a system of two inequalities with two unknowns, k and Fs . From (4.4) and (4.5) we can easily see that F s should be in the range 2FH 2FL ≤ Fs ≤ k k−1

(4.6)

To determine the integer k we rewrite (4.4) and (4.5) as follows: 1 k ≤ Fs 2FH

(4.7)

(k − 1)Fs ≤ 2FH − 2B

(4.8)

By multiplying (4.7) and (4.8) by sides and solving the resulting inequality for k we obtain FH kmax ≤ (4.9) B The maximum value of integer k is the number of bands that we can fit in the range from 0 to FH , that is  FH (4.10) kmax = B where b denotes the integer part of b. The minimum sampling rate required to avoid aliasing is Fsmax = 2FH /kmax . Therefore, the range of acceptable uniform sampling rates is determined by 2FH 2FL ≤ Fs ≤ k k−1

(4.11)

425

Sampling and Reconstruction of Signals

where k is an integer number given by  1≤k≤

FH B

(4.12)

As long as there is no aliasing, reconstruction is done using (4.2) and (4.3), which are valid for both integer and arbitrary band positioning. Choosing a Sampling Frequency. To appreciate the implications of conditions (4.11) and (4.12), we depict them graphically in Figure 4.3, as suggested by Vaughan et al. (1991). The plot shows the sampling frequency, normalized by B , as a function of the band position, FH /B . This is facilitated by rewriting (4.11) as follows:

Fs 2 FH 2 FH ≤ ≤ −1 (4.13) k B B k−1 B The shaded areas represent sampling rates that result in aliasing. The allowed range of sampling frequencies is inside the white wedges. For k = 1, we obtain 2FH ≤ Fs ≤ ∞, which is the sampling theorem for lowpass signals. Each wedge in the plot corresponds to a different value of k. To determine the allowed sampling frequencies, for a given FH and B , we draw a vertical line at the point determined by FH /B . The segments of the line within the allowed areas represent permissible sampling rates. We note that the theoretically minimum sampling frequency Fs = 2B , corresponding to integer band positioning,

H

s

F

st r qui

4

k=3

=

ate:

5

k=2 F

k=1

Fs = 2F H

6

Ny

Fs B 3

2 Forbidden (aliasing producing) region 1

1

2

3

4

5

6

7

8

9

10

FH B Allowed (white) and forbidden (shaded) sampling frequency regions for bandpass signals. The minimum sampling frequency Fs = 2B , which corresponds to the corners of the alias-free wedges, is possible for integerpositioned bands only.

Figure 4.3

426

Sampling and Reconstruction of Signals

occurs at the tips of the wedges. Therefore, any small variation of the sampling rate or the carrier frequency of the signal will move the sampling frequency into the forbidden area. A practical solution is to sample at a higher sampling rate, which is equivalent to augmenting the signal band with a guard band B = BL + BH . The augmented band locations and bandwidth are given by FL = FL − BL

(4.14)

FH = FH + BH

(4.15)

B = B + B

(4.16)

The lower-order wedge and the corresponding range of allowed sampling are given by  2FH 2FL FH ≤ F ≤ where k = (4.17) s k k −1 B The k th wedge with the guard bands and the sampling frequency tolerances are illustrated in Figure 4.4. The allowable range of sampling rates is divided into values above and below the practical operating points as Fs =

2FL 2FH = FsL + FsH − k − 1 k

(4.18)

From the shaded orthogonal triangles in Figure 4.4, we obtain BL =

k − 1 FsH 2

(4.19)

BH =

k FsL 2

(4.20)

which shows that symmetric guard bands lead to asymmetric sampling rate tolerance. 2 F´L Fs = k´−1 B B

∆FsH B ∆FsL B

Sampling frequency tolerance

Practical operating point

2 Fs = k´ B

F´B B

Figure 4.4

Illustration of the relationship between the size of guard bands and allowed sampling frequency deviations from its nominal value for the k th wedge.

∆BH ∆BL B B Guard-band widths

427

Sampling and Reconstruction of Signals

If we choose the practical operating point at the vertical midpoint of the wedge, the sampling rate is

2FL 1 2FH + (4.21) Fs = 2 k k −1 Since, by construction, FsL = FsH = Fs /2, the guard bands become BL =

k − 1 Fs 4

(4.22)

BH =

k Fs 4

(4.23)

We next provide an example that illustrates the use of this approach. EXAMPLE 4.1 Suppose we are given a bandpass signal with B = 25 kHz and FL = 10,702.5 kHz. From (4.10) the maximum wedge index is kmax = FH /B = 429 This yields the theoretically minimum sampling frequency Fs =

2FH = 50.0117 kHz kmax

To avoid potential aliasing due to hardware imperfections, we wish to use two guard bands of BL = 2.5 kHz and BH = 2.5 kHz on each side of the signal band. The effective bandwidth of the signal becomes B = B + BL + BH = 30 kHz. In addition, FL = FL − BL = 10,700 kHz and FH = FH + BH = 10,730 kHz. From (4.17), the maximum wedge index is kmax = FH /B = 357

Substitution of kmax into the inequality in (4.17) provides the range of acceptable sampling frequencies 60.1120 kHz ≤ Fs ≤ 60.1124 kHz

A detailed analysis on how to choose in practice the sampling rate for bandpass signals is provided by Vaughan et al. (1991) and Qi et al. (1996).

4.2

Interleaved or Nonuniform Second-Order Sampling

Suppose that we sample a continuous-time signal xa (t) with sampling rate Fi = 1/Ti at time instants t = nTi + i , where i is a fixed time offset. Using the sequence xi (nTi ) = xa (nTi + i ),

428

−∞ < n < ∞

(4.24)

Sampling and Reconstruction of Signals

and a reconstruction function ga(i) (t) we generate a continuous-time signal ya(i) (t) =

∞ 

xi (nTi )ga(i) (t − nTi − i )

(4.25)

n=−∞

The Fourier transform of ya(i) (t) is given by Ya(i) (F ) =

∞ 

−j 2πF (nTt +i ) xi (nTi )G(i) a (F )e

(4.26)

n=−∞ −j 2πFi = G(i) a (F )Xi (F )e

(4.27)

where Xi (F ) is the Fourier transform of xi (nTi ). From the sampling theorem (1.14), the Fourier transform of xi (nTi ) can be expressed in terms of the Fourier transform Xa (F )ej 2π Fi of xa (t + i ) as

∞ 1  k j 2π(F − Tk )i i Xa F − e Xi (F ) = Ti Ti

(4.28)

k=−∞

Substitution of (4.28) into (4.27) yields Ya(i) (F ) = G(i) a (F )

∞ 1  k −j 2π Tk i i Xa F − e Ti Ti

(4.29)

k=−∞

If we repeat the sampling process (4.24) for i = 1, 2, . . . , p, we obtain p interleaved uniformly sampled sequences xi (nTi ), −∞ < n < ∞. The sum of the p reconstructed signals is given by ya (t) =

p 

ya(i) (t)

(4.30)

i=1

Using (4.29) and (4.30), the Fourier transform of ya (t) can be expressed as Ya (F ) =

p 

(i) G(i) a (F )V (F )

(4.31)

i=1

where V

(i)

∞ 1  k −j 2π Tk i i (F ) = Xa F − e Ti Ti

(4.32)

k=−∞

We will focus on the most commonly used second-order sampling, defined by p = 2, 1 = 0, 2 = , T1 = T2 =

1 =T B

(4.33)

In this case, which is illustrated in Figure 4.5, relations (4.31) and (4.32) yield

429

Sampling and Reconstruction of Signals

xa(t) x1(nT)

x2(nT)

t t = nT

t = nT+∆ (a)

xa(t)

Ideal A/D

x1(nT)

Ideal D/A Ga(1)(F)

ya(1)(t)

ya(t)

t = nT

Ideal A/D

x2(nT)

Ideal D/A Ga(2)(F)

ya(2)(t)

t = nT+∆ (b) Illustration of second-order bandpass sampling: (a) interleaved sampled sequences (b) second-order sampling and reconstruction system.

Figure 4.5

Ya (F ) = BG(1) a (F )

∞ 

Xa (F − kB) + BG(2) a (F )

k=−∞

where

∞ 

γ k Xa (F − kB)

(4.34)

k=−∞

γ = e−j 2πB

(4.35)

To understand the nature of (4.34) we first split the spectrum X (F a ) into a “positive” band and a “negative” band as follows:   Xa (F ), F ≥ 0 Xa (F ), F ≤ 0 + − (4.36) Xa (F ) = , Xa (F ) = 0, F 0 Then, we plot the repeated replicas of Xa (F −kB) and γ k Xa (F −kB) as four separate components, as illustrated in Figure 4.6. We note that because each individual component has bandwidth B and sampling rate Fs = 1/B , its repeated copies fill the entire frequency axis without overlapping, that is, without aliasing. However, when we combine them, the negative bands cause aliasing to the positive bands, and vice versa.

430

Sampling and Reconstruction of Signals −

Xa(F)

Xa+(F)

Xa(F) 1

−FL − B

−FL

0

FL + B

FL

F

Σk Xa+(F − kB) k = −m − 1

k = −m

...

...

k=0

0

mB

F

Σk Xa−(F − kB) ...

k=0

...

k=m

k=m+1

0

F

mB

Σk γkXa+(F − kB) k = −m − 1

k = −m

...

...

k=0

0

F

Σk γ−kXa−(F − kB) k=0

...

... 0

Figure 4.6

k=m

k=m+1

−FL + mB

F

Illustration of aliasing in second-order bandpass sampling.

(2) We want to determine the interpolation functions G(1) a (F ), Ga (F ), and the time offset  so that Ya (F ) = Xa (F ). From Figure 4.6 we see that the first requirement is (2) G(1) a (F ) = Ga (F ) = 0, for |F | < FL and |F | > FL + B

(4.37)

(2) To determine G(1) a (F ) and Ga (F ) for FL ≤ |F | ≤ FL + B , we see from Figure 4.6 that only the components with k = ±m and k = ±(m + 1), where

2FL (4.38) m= B

is the smallest integer greater or equal to 2FL /B , overlap with the original spectrum.

431

Sampling and Reconstruction of Signals

In the region FL ≤ F ≤ −FL + mB , equation (4.34) becomes   (2) + Ya+ (F ) = BG(1) a (F ) + BGa (F ) Xa (F )

(Signal component)

  m (2) + BG(1) (F ) + Bγ G (F ) Xa+ (F − mB) a a

(Aliasing component)

The conditions that assure perfect reconstruction Ya+ (F ) = Xa+ (F ) are given by (2) BG(1) a (F ) + BGa (F ) = 1

(4.39)

m (2) BG(1) a (F ) + Bγ Ga (F ) = 0

(4.40)

Solving this system of equations yields the solution G(1) a (F ) =

1 1 , B 1 − γ −m

G(2) a (F ) =

1 1 B 1 − γm

(4.41)

which exists for all  such that γ ±m = e∓j 2πmB = 1. In the region −FL + mB ≤ F ≤ FL + B , equation (4.34) becomes   (2) (F ) + BG (F ) Xa+ (F ) Ya+ (F ) = BG(1) a a   m+1 (2) + BG(1) (F ) + Bγ G (F ) Xa+ (F − (m + 1)B) a a The conditions that assure perfect reconstruction Ya+ (F ) = Xa+ (F ) are given by (2) BG(1) a (F ) + BGa (F ) = 1

(4.42)

m+1 (2) BG(1) Ga (F ) = 0 a (F ) + Bγ

(4.43)

Solving this system of equations yields the solution G(1) a (F ) =

1 1 , B 1 − γ −(m+1)

G(2) a (F ) =

1 1 B 1 − γ m+1

which exists for all  such that γ ±(m+1) = e∓j 2π(m+1)B = 1.

432

(4.44)

Sampling and Reconstruction of Signals

The reconstruction functions in the frequency range −(FL + B) ≤ F ≤ −FL can be obtained in a similar manner. The formulas are given by (4.41) and (4.44) if we replace m by −m and m + 1 by −(m + 1). The function G(1) a (F ) has the bandpass (2) response shown in Figure 4.7. A similar plot for G a (F ) reveals that (1) G(2) a (F ) = Ga (−F )

(4.45)

which implies that ga(2) (t) = ga(1) (−t). Therefore, for simplicity, we adopt the notation ga (t) = ga(1) (t) and express the reconstruction formula (4.30) as follows xa (t) =

∞ 

xa

n=−∞

n

 n    n n ga t − + xa +  ga −t + +  B B B B

(4.46)

Taking the inverse Fourier transform of the function shown in Figure 4.7, we can show (see Problem 7) that the interpolation function is given by ga (t) = a(t) + b(t)

(4.47)

a(t) =

cos[2π(mB − FL )t − π mB] − cos(2π FL t − π mB) 2π Bt sin(π mB)

b(t) =

cos[2π(FL + B)t − π(m + 1)B] − cos[2π(mB − FL )t − π(m + 1)B] 2π Bt sin[π(m + 1)B]

(4.48)

(4.49) We can see that ga (0) = 1, ga (n/B) = 0 for n = 0, and ga (n/B ± ) = 0 for n = 0, ±1, ±2, . . ., as expected for any interpolation function. We have shown that a bandpass signal xa (t) with frequencies in the range FL ≤ |F | ≤ FL +B can be perfectly reconstructed from two interleaved uniformly sampled sequences xa (n/B) and xa (n/B + ), −∞ < n < ∞, using the interpolation formula (4.46) with an average rate F s = 2B samples/second without any restrictions on the band location. The time offset  cannot take values that may cause the interpolation function to take infinite values. This second-order sampling theorem was introduced by Kohlenberg (1953). The general pth-order sampling case (p > 2) is discussed by Coulson (1995). 1 1 − γ m+1

B −(FL + B)

BGa(1)(F) 1 1−γm

1 1 − γ −m

A FL − mB −FL

A 0

FL

1 1 − γ −(m+1)

B

−FL + mB

FL + B

F

Frequency domain characterization of the bandpass interpolation function for second-order sampling.

Figure 4.7

433

Sampling and Reconstruction of Signals

Some useful simplifications occur when m = 2FL /B , that is, for integer band positioning (Linden 1959, Vaughan et al. 1991). In this case, the region A becomes zero, which implies that a(t) = 0. Therefore, we have ga (t) = b(t). There are two cases of special interest. For low pass signals, FL = 0 and m = 0, and the interpolation function becomes gLP (t) =

cos(2π Bt − π B) − cos(π B) 2π Bt sin(π B)

(4.50)

The additional constraint  = 1/2B , which results in uniform sampling rate, yields the well-known sine interpolation function gLP (t) = sin(2π Bt)/2π Bt . For bandpass signals with FL = mB/2 we can choose the time offset  such that γ ±(m+1) = −1. This requirement is satisfied if =

1 k 2k + 1 = + , 2B(m + 1) 4Fc 2Fc

k = 0, ±1, ±2, . . .

(4.51)

where Fc = FL + B/2 = B(m + 1)/2 is the center frequency of the band. In this case, the interpolation function is specified by GQ (F ) = 1/2 in the range mB/2 ≤ |F | ≤ (m + 1)B/2 and GQ (F ) = 0 elsewhere. Taking the inverse Fourier transform, we obtain sin π Bt gQ (t) = cos 2π Fc t (4.52) π Bt which is a special case of (4.47)–(4.49). This special case is known as direct quadrature sampling because the in-phase and quadrature components are obtained explicitly from the bandpass signal (see Section 4.3). Finally, we note that it is possible to sample a bandpass signal, and then to reconstruct the discrete-time signal at a band position other than the original. This spectral relocation or frequency shifting of the bandpass signal is most commonly done using direct quadrature sampling [Coulson et al. (1994)]. The significance of this approach is that it can be implemented using digital signal processing.

4.3

Bandpass Signal Representations

The main cause of complications in the sampling of a real bandpass signal xa (t) is the presence of two separate spectral bands in the frequency regions −(FL + B) ≤ F ≤ −FL and FL ≤ F ≤ FL + B . Since xa (t) is real, the negative and positive frequencies in its spectrum are related by Xa (−F ) = Xa∗ (F )

(4.53)

Therefore, the signal can be completely specified by one half of the spectrum. We next exploit this idea to introduce simplified representations for bandpass signals. We start with the identity cos 2π Fc t =

434

1 j 2πFc t 1 −j 2πFc t + e e 2 2

(4.54)

Sampling and Reconstruction of Signals

which represents the real signal cos 2π Fc t by two spectral lines of magnitude 1/2, one at F = Fc and the other at F = −Fc . Equivalently, we have the identity  cos 2π Fc t = 2

1 j 2πFc t e 2

 (4.55)

which represents the real signal as the real part of a complex signal. In terms of the spectrum, we now specify the real signal cos 2π Fc t by the positive part of its spectrum, that is, the spectral line at F = Fc . The amplitude of the positive frequencies is doubled to compensate for the omission of the negative frequencies. The extension to signals with continuous spectra is straightforward. Indeed, the integral of the inverse Fourier transform of xa (t) can be split into two parts as 



xa (t) =

 Xa (F )ej 2πF t dF +

0

Xa (F )ej 2πF t dF

−∞

0

(4.56)

Changing the variable in the second integral from F to −F and using (4.53) yields  xa (t) =







Xa (F )ej 2πF t dF +

0

0

Xa∗ (F )e−j 2πF t dF

(4.57)

The last equation can be equivalently written as 





xa (t) = 

2Xa (F )e

j 2πF t

dF

= {ψa (t)}

(4.58)

0

where the complex signal  ψa (t) =



2Xa (F )ej 2πF t dF

(4.59)

0

is known as the analytic signal or the pre-envelope of xa (t). The spectrum of the analytic signal can be expressed in terms of the unit step function Va (F ) as follows: 

a (F ) = 2Xa (F )Va (F ) =

2Xa (F ), F > 0 0, F 0 HQ (F ) = hQ (t)e−j 2πF t dt = (4.65) j, F 0 F B at a sampling rate Fs = 1/T ≥ 2B , that is, x(n) = xa (nT ). Therefore, the spectrum X(F ) of x(n) is given by ∞ 1  X(F ) = Xa (F − kFs ) T

(5.2)

k=−∞

We next sample xa (t) at time instants t = nDT , that is, with a sampling rate Fs /D . The spectrum of the sequence xd (n) = xa (nDT ) is provided by

∞ 1  Fs Xd (F ) = Xa F − k DT D

(5.3)

k=−∞

This process is illustrated in Figure 5.1 for D = 2 and D = 4. We can easily see from Figure 5.1( c) that the spectrum Xd (F ) can be expressed in terms of the periodic spectrum X(F ) as

D−1 1  Fs X F −k Xd (F ) = (5.4) D D k=0

To avoid aliasing, the sampling rate should satisfy the condition Fs /D ≥ 2B . If the sampling frequency Fs is fixed, we can avoid aliasing by reducing the bandwidth of x(n) to (Fs /2)/D. In terms of the normalized frequency variables, we can avoid aliasing if the highest frequency fmax or ωmax in x(n) satisfies the conditions fmax ≤

1 fs = 2D 2

or

ωmax ≤

π ωs = D 2

(5.5)

439

Sampling and Reconstruction of Signals

xa(t )

Xa(F)

(a)

0 x(n) = xa(nT )

(b)

0

t T=

T

0

1 Fs

1 T t

−Fs

X (F )

−2Fs

2T

−Fs

0

x(n) = xa(n4T )

(d)

Figure 5.1

4T

t

−3Fs −2Fs −Fs

Fs =

0

Fs

1 2T

2Fs F

Fs Xd (F)

1 4T 0

1 T

Fs F Xd (F)

1 2T 0

Fs =

0

x(n) = xa(n2T )

(c)

F

1 Fs = 4T

2Fs 3Fs

F

Illustration of discrete-time signal sampling in the frequency domain.

In continuous-time sampling the continuous-time spectrum Xa (F ) is repeated an infinite number of times to create a periodic spectrum covering the infinite frequency range. In discrete-time sampling the periodic spectrum X(F ) is repeated D times to cover one period of the periodic frequency domain. To reconstruct the original sequence x(n) from the sampled sequence xd (n), we start with the ideal interpolation formula ∞  sin π (t − mDT ) xa (t) = xd (m) πDT (5.6) m=−∞ DT (t − mDT ) which reconstructs xa (t) assuming that Fs /D ≥ 2B . Since x(n) = xa (nT ), substitution into (5.6) yields ∞  sin π (n − mD) x(n) = xd (m) πD (5.7) m=−∞ D (n − mD) This is not a practical interpolator, since the sin(x)/x function is infinite in extent. In practice, we use a finite summation from m = −L to m = L. The quality of this approximation improves with increasing L. The Fourier transform of the ideal bandlimited interpolation sequence in (5.7) is  sin(π/D)n F D, |ω| ≤ π/D gBL (n) = D ←→ GBL (ω) = (5.8) 0, π/D < |ω| ≤ π πn Therefore, the ideal discrete-time interpolator has an ideal lowpass frequency characteristic.

440

Sampling and Reconstruction of Signals

x(m − 1)glin(t − mTd + Td )

x(m)

x(m)glin(t − mTd )

xlin(t) x(m − l)

Figure 5.2

Illustration of continuous-time linear interpolation.

(m − l)Td

C B t A mTd

To understand the process of discrete-time interpolation, we will analyze the widely used linear interpolation. For simplicity we use the notation Td = DT for the sampling period of xd (m) = xa (mTd ). The value of xa (t) at a time instant between mTd and (m + 1)Td is obtained by raising a vertical line from t to the line segment connecting the samples xd (mTd ) and xd (mTd + Td ), as shown in Figure 5.2. The interpolated value is given by xlin (t) = x(m − 1) +

x(m) − x(m − 1) [t − (m − 1)Td ], (m − 1)Td ≤ t ≤ mTd Td

which can be rearranged as follows:     t − (m − 1)Td mTd − t) x(m − 1) + 1 − x(m) xlin (t) = 1 − Td Td

(5.9)

(5.10)

To put (5.10) in the form of the general reconstruction formula xlin (t) =

∞ 

x(m)glin (t − mTd )

(5.11)

m=−∞

we note that we always have t − (m − 1)Td = |t − (m − 1)Td | and mTd − t = |t − mTd | because (m−1)Td ≤ t ≤ mTd . Therefore, we can express (5.10) in the form (5.11) if we define  |t| (5.12) glin (t) = 1 − Td , |t| ≤ Td 0, |t| > Td The discrete-time interpolation formulas are obtained by replacing t by nT in (5.11) and (5.12). Since T d = DT , we obtain ∞ 

xlin (n) =

x(m)glin (n − mD)

(5.13)

m=−∞

where

 glin (n) =

|n| 1− D , 0,

|n| ≤ D |n| > D

(5.14)

As expected from any interpolation function, glin (0) = 1 and glin (n) = 0 for n = ±D, ±2D, . . .. The performance of the linear interpolator can be assessed by comparing its Fourier transform   1 sin(ωD/2) 2 Glin (ω) = (5.15) D sin(ω/2)

441

Sampling and Reconstruction of Signals

L GBL(ω) L=5 Glin(ω)

Figure 5.3

Frequency response of ideal and linear discrete-time interpolators.

−π − 4π 5

− 2π − π 5 5

0

π 5

2π 5

4π 5

π

ω

to that of the ideal interpolator (5.8). This is illustrated in Figure 5.3 which shows that the linear interpolator has good performance only when the spectrum of the interpolated signal is negligible for |ω| > π/D , that is, when the original continuoustime signal has been oversampled. Equations (5.11) and (5.13) resemble a convolution summation; however, they are not convolutions. This is illustrated in Figure 5.4 which shows the computation of interpolated samples x(nT ) and x((n + 1)T ) for D = 5. We note that only a subset of the coefficients of the linear interpolator is used in each case. Basically, we decompose glin (n) into D components and we use one at a time periodically to compute the interpolated values. This is essentially the idea behind polyphase filter by inserting (D − 1) zero structures. However, if we create a new sequence x(n) ˜ samples between successive samples of xd (m), we can compute x(n) using the convolution x(n) =

∞ 

x(k)g ˜ lin (n − k)

(5.16)

k=−∞

at the expense of unnecessary computations involving zero values. A more efficient implementation can be obtained using equation (5.13). Sampling and interpolation of a discrete-time signal essentially corresponds to a change of its sampling rate by an integer factor. The subject of sampling rate conversion, which is very important in practical applications. Extensive discussion is beyond the scope of this chapter.

5.2

Representation and Sampling of Bandpass Discrete-Time Signals

The bandpass representations of continuous-time signals, discussed in Section 4.3, can be adapted for discrete-time signals with some simple modifications that take into consideration the periodic nature of discrete-time spectra. Since we cannot require that the discrete-time Fourier transform is zero for ω < 0 without violating its periodicity, we define the analytic signal ψ(n) of a bandpass sequence x(n) by 

(ω) =

2X(ω), 0 ≤ ω < π 0, −π ≤ ω < 0

(5.17)

where X(ω) and (ω) are the Fourier transforms of x(n) and ψ(n), respectively.

442

Sampling and Reconstruction of Signals

x(m − 1)

(m − 1)Td

glin(nT − mTd )

nT

x(m)

t

mTd xlin(n)

t

nT

~ x(n)

glin(n − k)

(n − 1)T nT (n + 1)T Figure 5.4

t

Illustration of linear interpolation as a linear filtering process.

The ideal discrete-time Hilbert transformer, defined by  −j, 0 < ω < π H (ω) = j, −π < ω < 0

(5.18)

is a 90-degree phase shifter as in the continuous-time case. We can easily show that

where

ˆ

(ω) = X(ω) + j X(ω)

(5.19)

ˆ X(ω) = H (ω)X(ω)

(5.20)

To compute the analytic signal in the time domain, we need the impulse response of the Hilbert transformer. It is obtained by  0  π 1 1 j ωn (5.21) j e dω − j ej ωn dω h(n) = 2π −π 2π 0

443

Sampling and Reconstruction of Signals

which yields  h(n) =



2 2 sin (π n/2) , π n 0,

n = 0 = n=0

0, 2 πn,

n = even n = odd

(5.22)

The sequence h(n) is nonzero for n < 0 and not absolutely summable; thus, the ideal Hilbert transformer is noncausal and unstable. The impulse response and the frequency response of the ideal Hilbert transformer are illustrated in Figure 5.5. As in the continuous-time case the Hilbert transform x(n) ˆ of a sequence x(n) provides the imaginary part of its analytic signal representation, that is, ψ(n) = x(n) + j x(n) ˆ

(5.23)

The complex envelope, quadrature, and envelope/phase representations are obtained by the corresponding formulas for continuous-time signals by replacing t by nT in all relevant equations. Given a bandpass sequence x(n), 0 < ωL ≤ |ω| ≤ ωL + w with normalized bandwidth w = 2πB/Fs , we can derive equivalent complex envelope or in-phase and quadrature lowpass representations that can be sampled at a rate fs = 1/D h[n]

−7

−5

−5

−3

−1 0

1

2

3

4

5

6

7

8

(a) H(ω) j

π

−π

−j

(b)

Impulse response (a) and frequency response (b) of the discrete-time Hilbert transformer.

Figure 5.5

444

n

Sampling and Reconstruction of Signals

compatible with the bandwidth w. If ωL = (k − 1)π/D and w = π/D the sequence x(n) can be sampled directly without any aliasing. In many radar and communication systems applications it is necessary to process a bandpass signal xa (t), FL ≤ |F | ≤ FL + B in lowpass form. Conventional techniques employ two quadrature analog channels and two A/D converters following the two lowpass filters in Figure 4.8. A more up-to-date approach is to uniformly sample the analog signal and then obtain the quadrature representation using digital quadrature demodulation, that is, a discrete-time implementation of the first part of Figure 4.8. A similar approach can be used to digitally generate single sideband signals for communications applications (Frerking 1994).

6

Oversampling A/D and D/A Converters In this section we treat oversampling A/D and D/A converters.

6.1

Oversampling A/D Converters

The basic idea in oversampling A/D converters is to increase the sampling rate of the signal to the point where a low-resolution quantizer suffices. By oversampling, we can reduce the dynamic range of the signal values between successive samples and thus reduce the resolution requirements on the quantizer. As we have observed in the preceding section, the variance of the quantization error in A/D conversion is σe2 = 2 /12, where  = R/2b+1 . Since the dynamic range of the signal, which is proportional to its standard deviation σx , should match the range R of the quantizer, it follows that  is proportional to σx . Hence for a given number of bits, the power of the quantization noise is proportional to the variance of the signal to be quantized. Consequently, for a given fixed SQNR, a reduction in the variance of the signal to be quantized allows us to reduce the number of bits in the quantizer. The basic idea for reducing the dynamic range leads us to consider differential quantization. To illustrate this point, let us evaluate the variance of the difference between two successive signal samples. Thus we have d(n) = x(n) − x(n − 1)

(6.1)

The variance of d(n) is σd2 = E[d 2 (n)] = E{[x(n) − x(n − 1)]2 } = E[x 2 (n)] − 2E[x(n)x(n − 1)] + E[x 2 (n − 1)]

(6.2)

= 2σx2 [1 − γxx (1)] where γxx (1) is the value of the autocorrelation sequence γxx (m) of x(n) evaluated at m = 1. If γxx (1) > 0.5, we observe that σd2 < σx2 . Under this condition, it is better to quantize the difference d(n) and to recover x(n) from the quantized values {dq (n)}. To obtain a high correlation between successive samples of the signal, we require that the sampling rate be significantly higher than the Nyquist rate.

445

Sampling and Reconstruction of Signals

An even better approach is to quantize the difference d(n) = x(n) − ax(n − 1)

(6.3)

where a is a parameter selected to minimize the variance in d(n). This leads to the result (see Problem 16) that the optimum choice of a is a=

γxx (1) γxx (1) = γxx (0) σx2

and σd2 = σx2 [1 − a 2 ]

(6.4)

In this case, σd2 < σx2 , since 0 ≤ a ≤ 1. The quantity ax(n − 1) is called a first-order predictor of x(n). Figure 6.1 shows a more general differential predictive signal quantizer system. This system is used in speech encoding and transmission over telephone channels and is known as differential pulse code modulation (DPCM). The goal of the predictor is to provide an estimate x(n) ˆ of x(n) from a linear combination of past values of x(n), so as to reduce the dynamic range of the difference signal d(n) = x(n) − x(n). ˆ Thus a predictor of order p has the form x(n) ˆ =

p 

ak x(n − k)

(6.5)

k=1

The use of the feedback loop around the quantizer as shown in Fig. 6.1 is necessary to avoid the accumulation of quantization errors at the decoder. In this configuration, the error e(n) = d(n) − dq (n) is e(n) = d(n) − dq (n) = x(n) − x(n) ˆ − dq (n) = x(n) − xq (n) Thus the error in the reconstructed quantized signal xq (n) is equal to the quantization error for the sample d(n). The decoder for DPCM that reconstructs the signal from the quantized values is also shown in Fig. 6.1. The simplest form of differential predictive quantization is called delta modulation (DM). In DM, the quantizer is a simple 1-bit (two-level) quantizer and the x(n)

d(n) + −

xq(n)

dq(n) +

Q[ ] ^ x(n)

PR

Coder

xq(n)

^ x(n)

+

PR

Decoder

Encoder and decoder for differential predictive signal quantizer system.

Figure 6.1

446

Sampling and Reconstruction of Signals

predictor is a first-order predictor, as shown in Fig. 6.2(a). Basically, DM provides a staircase approximation of the input signal. At every sampling instant, the sign of the difference between the input sample x(n) and its most recent staircase approximation x(n) ˆ = axq (n − 1) is determined, and then the staircase signal is updated by a step  in the direction of the difference. From Fig. 6.2(a) we observe that xq (n) = axq (n − 1) + dq (n)

(6.6)

which is the discrete-time equivalent of an analog integrator. If a = 1, we have an ideal accumulator (integrator) whereas the choice a < 1 results in a “leaky integrax(n)

d(n)

dq(n)

+1

+ − a

^ x(n)

z−1

xq(n)

xq(n)

+

−1 ^ x(n)

a

+

Coder

z−1

Decoder (a)

x(n)

Slope-overload distortion

Granular noise

x(n − 1) Step-size ∆ xa(t) ^ x(n)

Time T=

1 Fs (b) T

T +1

x(t)

+

d(t)

−1

+1 −1

Clock



^ x(t)

LPF

^ x(t)



^ x(t)



Integrator (c)

Figure 6.2

Delta modulation system and two types of quantization errors.

447

Sampling and Reconstruction of Signals

tor.” Figure 6.2(c) shows an analog model that illustrates the basic principle for the practical implementation of a DM system. The analog lowpass filter is necessary for the rejection of out-of-band components in the frequency range between B and Fs /2, since Fs >> B due to oversampling. The crosshatched areas in Fig. 6.2(b) illustrate two types of quantization error in DM, slope-overload distortion and granular noise. Since the maximum slope /T in x(n) is limited by the step size, slope-overload distortion can be avoided if max |dx(t)/dt| ≤ /T . The granular noise occurs when the DM tracks a relatively flat (slowly changing) input signal. We note that increasing  reduces overload distortion but increases the granular noise, and vice versa. One way to reduce these two types of distortion is to use an integrator in front of the DM, as shown in Fig. 6.3(a). This has two effects. First, it emphasizes the low frequencies of x(t) and increases the correlation of the signal into the DM input. Second, it simplifies the DM decoder because the differentiator (inverse system) required at the decoder is canceled by the DM integrator. Hence the decoder is simply a lowpass filter, as shown in Fig. 6.3(a). Furthermore, the two integrators at the encoder can be replaced by a single integrator placed before the comparator, as shown in Fig. 6.3(b). This system is known as sigma-delta modulation (SDM). SDM is an ideal candidate for A/D conversion. Such a converter takes advantage of the high sampling rate and spreads the quantization noise across the band up to Fs /2. Since Fs >> B , the noise in the signal-free band B ≤ F ≤ Fs /2 can be Clock x(t)



Analog LPF

+1

+

−1



∫ Coder

Decoder (a) Clock

x(t) +



+1 −1

Analog LPF



Coder

Decoder (b)

Figure 6.3

448

Sigma-delta modulation system.

Sampling and Reconstruction of Signals

H(z) x(n) +

e(n) d(n)

z−1

+

+

dq(n)



Figure 6.4

Discrete-time model of sigma-delta modulation.

removed by appropriate digital filtering. To illustrate this principle, let us consider the discrete-time model of SDM, shown in Fig. 6.4, where we have assumed that the comparator (1-bit quantizer) is modeled by an additive white noise source with variance σe2 = 2 /12. The integrator is modeled by the discrete-time system with system function z−1 H (z) = (6.7) 1 − z−1 The z-transform of the sequence {dq (n)} is Dq (z) =

H (z) 1 X(z) + E(z) 1 + H (z) 1 + H (z)

(6.8)

= Hs (z)X(z) + Hn (z)E(z) where Hs (z) and Hn (z) are the signal and noise system functions, respectively. A good SDM system has a flat frequency response Hs (ω) in the signal frequency band 0 ≤ F ≤ B . On the other hand, Hn (z) should have high attenuation in the frequency band 0 ≤ F ≤ B and low attenuation in the band B ≤ F ≤ Fs /2. For the first-order SDM system with the integrator specified by (6.7), we have Hs (z) = z−1 ,

Hn (z) = 1 − z−1

(6.9)

Thus Hs (z) does not distort the signal. The performance of the SDM system is therefore determined by the noise system function Hn (z), which has a magnitude frequency response    π F   (6.10) |Hn (F )| = 2 sin F  s

as shown in Fig. 6.5. The in-band quantization noise variance is given as  B |Hn (F )|2 Se (F ) dF σn2 = −B

(6.11)

where Se (F ) = σe2 /Fs is the power spectral density of the quantization noise. From this relationship we note that doubling Fs (increasing the sampling rate by a factor of 2), while keeping B fixed, reduces the power of the quantization noise by 3 dB. This result is true for any quantizer. However, additional reduction may be possible by properly choosing the filter H (z).

449

Sampling and Reconstruction of Signals

Sr(F)



Figure 6.5

Fs 2

σ2e /Fs

Hn(F)

−B

B

Fs 2

F

Frequency (magnitude) response of noise system function.

For the first-order SDM, it can be shown (see Problem 19) that for Fs >> 2B , the in-band quantization noise power is σn2

1 ≈ π 2 σe2 3



2B Fs

3 (6.12)

Note that doubling the sampling frequency reduces the noise power by 9 dB, of which 3 dB is due to the reduction in Se (F ) and 6 dB is due to the filter characteristic Hn (F ). An additional 6-dB reduction can be achieved by using a double integrator (see Problem 20). In summary, the noise power σn2 can be reduced by increasing the sampling rate to spread the quantization noise power over a larger frequency band (−Fs /2, Fs /2), and then shaping the noise power spectral density by means of an appropriate filter. Thus, SDM provides a 1-bit quantized signal at a sampling frequency Fs = 2IB , where the oversampling (interpolation) factor I determines the SNR of the SDM quantizer. Next, we explain how to convert this signal into a b-bit quantized signal at the Nyquist rate. First, we recall that the SDM decoder is an analog lowpass filter with a cutoff frequency B . The output of this filter is an approximation to the input signal x(t). Given the 1-bit signal dq (n) at sampling frequency Fs , we can obtain a signal xq (n) at a lower sampling frequency, say the Nyquist rate of 2B or somewhat faster, by resampling the output of the lowpass filter at the 2B rate. To avoid aliasing, we first filter out the out-of-band (B, Fs /2) noise by processing the wideband signal. The signal is then passed through the lowpass filter and resampled (down-sampled) at the lower rate. The down-sampling process is called decimation. For example, if the interpolation factor is I = 256, the A/D converter output can be obtained by averaging successive nonoverlapping blocks of 128 bits. This averaging would result in a digital signal with a range of values from zero to 256(b ≈ 8 bits) at the Nyquist rate. The averaging process also provides the required antialiasing filtering. Figure 6.6 illustrates the basic elements of an oversampling A/D converter. Oversampling A/D converters for voiceband (3-kHz) signals are currently fabricated

450

Sampling and Reconstruction of Signals

Digital section

Analog section

x(t)

Antialiasing filter

SDM

1-bit dq(n)

Digital LPF (decimator)

Fs

b>1

xq(n)

FN

SDM-to-PCM converter Digital section b-bit FN

Digital LPF (interpolator)

b-bit

Analog section

Digital SDM

Fs

1-bit Fs

Sampled data LPF

PCM-to-SDM converter

Figure 6.6

Smoothing filter

Antialiasing filters

Basic elements of an oversampling A/D converter.

as integrated circuits. Typically, they operate at a 2-MHz sampling rate, down-sample to 8 kHz, and provide 16-bit accuracy.

6.2

Oversampling D/A Converters

The elements of an oversampling D/A converter are shown in Fig. 6.7. As we observe, it is subdivided into a digital front end followed by an analog section. The digital section consists of an interpolator whose function is to increase the sampling rate by some factor I , which is followed by an SDM. The interpolator simply increases the digital sampling rate by inserting I −1 zeros between successive low rate samples. The resulting signal is then processed by a digital filter with cutoff frequency Fc = B/Fs in order to reject the images (replicas) of the input signal spectrum. This higher rate signal is fed to the SDM, which creates a noise-shaped 1-bit sample. Each 1-bit sample is fed to the 1-bit D/A, which provides the analog interface to the antialiasing and smoothing filters. The output analog filters have a passband of 0 ≤ F ≤ B hertz and serve to smooth the signal and to remove the quantization noise in the frequency band B ≤ F ≤ Fs /2. In effect, the oversampling D/A converter uses SDM with the roles of the analog and digital sections reversed compared to the A/D converter. In practice, oversampling D/A (and A/D) converters have many advantages over the more conventional D/A (and A/D) converters. First, the high sampling rate and Digital signal

Interpolation filter

Sigmadelta modulator

Digital section

Figure 6.7

Analog smoothing filter

1-bit D/A

Analog output

Analog section

Elements of an oversampling D/A converter.

451

Sampling and Reconstruction of Signals

the subsequent digital filtering minimize or remove the need for complex and expensive analog antialiasing filters. Furthermore, any analog noise introduced during the conversion phase is filtered out. Also, there is no need for S/H circuits. Oversampling SDM A/D and D/A converters are very robust with respect to variations in the analog circuit parameters, are inherently linear, and have low cost. This concludes our discussion of signal reconstruction based on simple interpolation techniques. The techniques that we have described are easily incorporated into the design of practical D/A converters for the reconstruction of analog signals from digital signals.

7

Summary and References The major focus of this chapter was on the sampling and reconstruction of signals. In particular, we treated the sampling of continuous-time signals and the subsequent operation of A/D conversion. These are necessary operations in the digital processing of analog signals, either on a general-purpose computer or on a custom-designed digital signal processor. The related issue of D/A conversion was also treated. In addition to the conventional A/D and D/A conversion techniques, we also described another type of A/D and D/A conversion, based on the principle of oversampling and a type of waveform encoding called sigma-delta modulation. Sigma-delta conversion technology is especially suitable for audio band signals due to their relatively small bandwidth (less than 20 kHz) and in some applications, the requirements for high fidelity. The sampling theorem was introduced by Nyquist (1928) and later popularized in the classic paper by Shannon (1949). D/A and A/D conversion techniques are treated in a book by Sheingold (1986). Oversampling A/D and D/A conversion has been treated in the technical literature. Specifically, we cite the work of Candy (1986), Candy et al. (1981) and Gray (1990).

Problems Given a continuous-time signal xa (t) with Xa (F ) = 0 for |F | > B determine the minimum sampling rate Fs for a signal ya (t) defined by (a) dxa (t)/dt (b) xa2 (t) (c) xa (2t) (d) xa (t) cos 6πBt and (e) xa (t) cos 7π Bt 2 The sampled sequence xa (nT ) is reconstructed using an ideal D/A with interpolation function ga (t) = A for |F | < Fc and zero otherwise to produce a continuous-time signal xˆ a (t).

1

(a) If the spectrum of the original signal xa (t) satisfies Xa (F ) = 0 for |F | > B , find the maximum value of T , and the values of Fc , and A such that xˆ a (t) = xa (t). (b) If X1 (F ) = 0 for |F | > B , X2 (F ) = 0 for |F | > 2B , and xa (t) = x1 (t)x2 (t), find the maximum value of T , and the values of Fc , and A such that xˆ a (t) = xa (t). (c) Repeat part (b) for xa (t) = x1 (t)x2 (t/2).

452

Sampling and Reconstruction of Signals

A continuous-time periodic signal with Fourier series coefficients ck = (1/2)|k| and period Tp = 0.1 sec passes through an ideal lowpass filter with cutoff frequency Fc = 102.5 Hz. The resulting signal ya (t) is sampled periodically with T = 0.005 sec. Determine the spectrum of the sequence y(n) = ya (nT ). 4 Repeat Example 1.2 for the signal xa (t) = te −t ua (t). 5 Consider the system in Figure 2.1. If Xa (F ) = 0 for |F | > Fs /2, determine the t frequency response H (ω) of the discrete-time system such that ya (t) = −∞ xa (τ )dτ . 6 Consider a signal xa (t) with spectrum Xa (F ) = 0 for 0 < F1 ≤ |F | ≤ F2 < ∞ and Xa (F ) = 0 otherwise. 3

(a) Determine the minimum sampling frequency required to sample xa (t) without aliasing. (b) Find the formula needed to reconstruct xa (t) from the samples xa (nT ), −∞ < n < ∞. 7 Prove the nonuniform second-order sampling interpolation formula described by equations (4.47)–(4.49). 8 A discrete-time sample-and-hold interpolator, by a factor I, repeats the last input sample (I − 1) times. (a) Determine the interpolation function gSH (n). (b) Determine the Fourier transform GSH (ω) of gSH (n). (c) Plot the magnitude and phase responses of the ideal interpolator, the linear interpolator, and the sample-and-hold interpolator for I = 5. 9 Time-domain sampling Consider the continuous-time signal  xa (t) =

e−j 2πF0 t , 0,

t ≥0 t B , show that

3 1 2 2 2B σn ≈ π2σe 3 Fs

456

Sampling and Reconstruction of Signals

20

Consider the second-order SDM model shown in Fig. P20. e(n) x(n) +

+

+



z−1

+

+

dq(n)

− z−1

z−1

Figure P20

(a) Determine the signal and noise system functions Hs (z) and Hn (z), respectively. (b) Plot the magnitude response for the noise system function and compare it with the one for the first-order SDM. Can you explain the 6-dB difference from these curves? (c) Show that the in-band quantization noise power σn2 is given approximately by

π σe2 2B 5 2 σn ≈ 5 Fs 21

which implies a 15-dB increase for every doubling of the sampling frequency. Figure P21 illustrates the basic idea for a lookup-table-based sinusoidal signal generator. The samples of one period of the signal

2π n , n = 0, 1, . . . , N − 1 x(n) = cos N are stored in memory. A digital sinusoidal signal is generated by stepping through the table and wrapping around at the end when the angle exceeds 2π . This can be done by using modulo-N addressing (i.e., using a “circular” buffer). Samples of x(n) are feeding the ideal D/A converter every T seconds. (a) Show that by changing Fs we can adjust the frequency F0 of the resulting analog sinusoid. (b) Suppose now that Fs = 1/T is fixed. How many distinct analog sinusoids can be generated using the given lookup table? Explain. x(0)

x(1)



DSP

x(N − 1)

Ideal D/A

Fs =

xa(t) = cos 2πF0 t

1 T

Figure P21

457

Sampling and Reconstruction of Signals

22 Suppose that we represent an analog bandpass filter by the frequency response H (F ) = C(F − Fc ) + C ∗ (−F − Fc ) where C(f ) is the frequency response of an equivalent lowpass filter, as shown in Fig. P22. (a) Show that the impulse response c(t) of the equivalent lowpass filter is related to the impulse response h(t) of the bandpass filter as follows: h(t) = 2[c(t)ej 2πFc t ] (b) Suppose that the bandpass system with frequency response H (F ) is excited by a bandpass signal of the form x(t) = [u(t)ej 2πFc t ] where u(t) is the equivalent lowpass signal. Show that the filter output may be expressed as y(t) = [v(t)ej 2πFc t ] where

 v(t) =

ω

c(τ )u(t − τ )dτ

(Hint: Use the frequency domain to prove this result.) C(F)

Figure P22

F −B

0

B

23 Consider the sinusoidal signal generator in Fig. P23, where both the stored sinusoidal data

2A x(n) = cos n , 0≤n≤N −1 N and the sampling frequency Fs = 1/T are fixed. An engineer wishing to produce a sinusoid with period 2N suggests that we use either zero-order or first-order (linear) interpolation to double the number of samples per period in the original sinusoid as illustrated in Fig. P23(a).

458

Sampling and Reconstruction of Signals

Interpolated values

x(n) 0

1

2

3

4

5

6

7

xt(n)

Insert zeros

H(z)

y(n)

(b)

Zero-order interpolation Interpolated values X(ω) 1

0

1

2

3

4

5

6

−π 3

0

π 3

(c)

Linear interpolation (a)

Figure P23

(a) Determine the signal sequences y(n) generated using zero-order interpolation and linear interpolation and then compute the total harmonic distortion (THD) in each case for N = 32, 64, 128. (b) Repeat part (a) assuming that all sample values are quantized to 8 bits. (c) Show that the interpolated signal sequences y(n) can be obtained by the system shown in Fig. P23 (b). The first module inserts one zero sample between successive samples of x(n). Determine the system H (z) and sketch its magnitude response for the zero-order interpolation and for the linear interpolation cases. Can you explain the difference in performance in terms of the frequency response functions? (d) Determine and sketch the spectra of the resulting sinusoids in each case both analytically [using the results in part (c)] and evaluating the DFT of the resulting signals. (e) Sketch the spectra of xi (n) and y(n), if x(n) has the spectrum shown in Fig. P23 (c) for both zero-order and linear interpolation. Can you suggest a better choice for H (z)?

459

Sampling and Reconstruction of Signals

24 Let xa (t) be a time-limited signal; that is, xa (t) = 0 for |t| > τ , with Fourier transform Xa (F ). The function Xa (F ) is sampled with sampling interval δF = 1/Ts . (a) Show that the function ∞ 

xp (t) =

xa (t − nTs )

n=−∞

can be expressed as a Fourier series with coefficients ck =

1 Xa (kδF ) Ts

(b) Show that Xa (F ) can be recovered from the samples Xa (kδF ), −∞ < k < ∞ if Ts ≥ 2τ . (c) Show that if Ts < 2τ , there is “time-domain aliasing” that prevents exact reconstruction of Xa (F ). (d) Show that if Ts ≥ 2τ , perfect reconstruction of Xa (F ) from the samples X(kδF ) is possible using the interpolation formula ∞ 

Xa (F ) =

Xa (kδF )

k=−∞

sin[(π/δF )(F − kδF )] (π/δF )(F − kδF )

Answers to Selected Problems 9

(a) Xa (F ) = (b) X(f ) =

10 Since 14

Fc + B2 B

∞ n=−∞

=

1 j 2π(F +F0 )

1 1−e−j 2π(F +F0 /Fs )

50+10 20

x 2 (n) =

1 2π

= 3 is an integer, then Fs = 2B = 40 Hz

π −π

|X(w)|2 dw =

Ea T

18 Let Pd denote the power spectral density of the quantization noise. Then  B Pn = −FsB Pd df = 2B P = σe2 Fs d Fs

σ2

σ 2F

σ 2F

x s x s SQNR = 10 log 10 σx2 = 10 log10 2BP = 10 log10 2BP + 10 log 10Fs d d e Thus, SQNR will increase by 3 dB if Fs is doubled.

20 Hs (z) = z−1 ; Hn (z) = (1 − z−1 )2

460

The Discrete Fourier Transform: Its Properties and Applications

Frequency analysis of discrete-time signals is usually and most conveniently performed on a digital signal processor, which may be a general-purpose digital computer or specially designed digital hardware. To perform frequency analysis on a discrete-time signal {x(n)}, we convert the time-domain sequence to an equivalent frequency-domain representation. We know that such a representation is given by the Fourier transform X(ω) of the sequence {x(n)}. However, X(ω) is a continuous function of frequency and therefore it is not a computationally convenient representation of the sequence {x(n)}. In this chapter we consider the representation of a sequence {x(n)} by samples of its spectrum X(ω). Such a frequency-domain representation leads to the discrete Fourier transform (DFT), which is a powerful computational tool for performing frequency analysis of discrete-time signals.

1

Frequency-Domain Sampling: The Discrete Fourier Transform Before we introduce the DFT, we consider the sampling of the Fourier transform of an aperiodic discrete-time sequence. Thus, we establish the relationship between the sampled Fourier transform and the DFT.

1.1

Frequency-Domain Sampling and Reconstruction of Discrete-Time Signals

We recall that aperiodic finite-energy signals have continuous spectra. Let us consider such an aperiodic discrete-time signal x(n) with Fourier transform X(ω) =

∞ 

x(n)e−j ωn

(1.1)

n=−∞

From Chapter 7 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

461

The Discrete Fourier Transform: Its Properties and Applications

X(ω)

X(kδω)

Figure 1.1

Frequency-domain sampling of the Fourier transform.

−π

0

kδω

π

 δω  2π

ω

Suppose that we sample X(ω) periodically in frequency at a spacing of δω radians between successive samples. Since X(ω) is periodic with period 2π , only samples in the fundamental frequency range are necessary. For convenience, we take N equidistant samples in the interval 0 ≤ ω < 2π with spacing δω = 2π/N , as shown in Fig. 1.1. First, we consider the selection of N , the number of samples in the frequency domain. If we evaluate (1.1) at ω = 2π k/N , we obtain 

 ∞  2π X k = x(n)e−j 2πkn/N , N n=−∞

k = 0, 1, . . . , N − 1

(1.2)

The summation in (1.2) can be subdivided into an infinite number of summations, where each sum contains N terms. Thus  2π k = X N 

+

−1 

··· +

x(n)e−j 2πkn/N +

n=−N 2N−1 

N−1 

x(n)e−j 2πkn/N

n=0

x(n)e−j 2πkn/N + · · ·

n=N

=

∞ 

lN +N−1

l=−∞

n=lN

x(n)e−j 2πkn/N

If we change the index in the inner summation from n to n − lN and interchange the order of the summation, we obtain the result  ∞   N−1   2π k = x(n − lN ) e−j 2πkn/N X N 

n=0

(1.3)

l=−∞

for k = 0, 1, 2, . . . , N − 1. The signal xp (n) =

∞  l=−∞

462

x(n − lN )

(1.4)

The Discrete Fourier Transform: Its Properties and Applications

obtained by the periodic repetition of x(n) every N samples, is clearly periodic with fundamental period N . Consequently, it can be expanded in a Fourier series as xp (n) =

N−1 

n = 0, 1, . . . , N − 1

ck ej 2πkn/N ,

(1.5)

k=0

with Fourier coefficients ck =

N−1 1  xp (n)e−j 2πkn/N , N

k = 0, 1, . . . , N − 1

(1.6)

n=0

Upon comparing (1.3) with (1.6), we conclude that   1 2π ck = X k , N N

k = 0, 1, . . . , N − 1

(1.7)

Therefore, xp (n) =

  N−1 1  2π X k ej 2πkn/N , N N

n = 0, 1, . . . , N − 1

(1.8)

k=0

The relationship in (1.8) provides the reconstruction of the periodic signal xp (n) from the samples of the spectrum X(ω). However, it does not imply that we can recover X(ω) or x(n) from the samples. To accomplish this, we need to consider the relationship between xp (n) and x(n). Since xp (n) is the periodic extension of x(n) as given by (1.4), it is clear that x(n) can be recovered from xp (n) if there is no aliasing in the time domain, that is, if x(n) is time-limited to less than the period N of xp (n). This situation is illustrated in Fig. 1.2, where without loss of generality, we consider a finite-duration sequence x(n), which is nonzero in the interval 0 ≤ n ≤ L − 1. We observe that when N ≥ L, x(n) = xp (n),

0≤n≤N −1

so that x(n) can be recovered from xp (n) without ambiguity. On the other hand, if N < L, it is not possible to recover x(n) from its periodic extension due to timedomain aliasing. Thus, we conclude that the spectrum of an aperiodic discrete-time signal with finite duration L can be exactly recovered from its samples at frequencies ωk = 2π k/N, if N ≥ L. The procedure is to compute xp (n), n = 0, 1, . . . , N − 1 from (1.8); then  xp (n), 0≤n≤N −1 (1.9) x(n) = 0, elsewhere and finally, X(ω) can be computed from (1.1).

463

The Discrete Fourier Transform: Its Properties and Applications

x(n)

n 0

L

xp(n) N>L

n 0

L

N

xp(n) N M without loss of generality.

497

The Discrete Fourier Transform: Its Properties and Applications

Overlap-save method. In this method the size of the input data blocks is N = L + M − 1 and the DFTs and IDFT are of length N . Each data block consists of the last M − 1 data points of the previous data block followed by L new data points to form a data sequence of length N = L + M − 1. An N -point DFT is computed for each data block. The impulse response of the FIR filter is increased in length by appending L − 1 zeros and an N -point DFT of the sequence is computed once and stored. The multiplication of the two N -point DFTs {H (k)} and {Xm (k)} for the mth block of data yields Yˆ m (k) = H (k)Xm (k),

k = 0, 1, . . . , N − 1

(3.7)

Then the N -point IDFT yields the result Yˆ m (n) = {yˆ m (0)yˆ m (1) · · · yˆ m (M − 1)yˆ m (M) · · · yˆ m (N − 1)}

(3.8)

Since the data record is of length N, the first M − 1 points of ym (n) are corrupted by aliasing and must be discarded. The last L points of ym (n) are exactly the same as the result from linear convolution and, as a consequence, yˆ m (n) = ym (n), n = M, M + 1, . . . , N − 1

(3.9)

To avoid loss of data due to aliasing, the last M − 1 points of each data record are saved and these points become the first M − 1 data points of the subsequent record, as indicated above. To begin the processing, the first M − 1 points of the first record are set to zero. Thus the blocks of data sequences are x1 (n) = {0, 0, . . . , 0, x(0), x(1), . . . , x(L − 1)}   

(3.10)

M−1 points

x2 (n) = {x(L − M + 1), . . . , x(L − 1), x(L), . . . , x(2L − 1)}       M−1data points from x1 (n)

L new data points

x3 (n) = {x(2L − M + 1), . . . , x(2L − 1), x(2L), . . . , x(3L − 1)}       M−1 data points from x2 (n)

(3.11)

(3.12)

L new data points

and so forth. The resulting data sequences from the IDFT are given by (3.8), where the first M − 1 points are discarded due to aliasing and the remaining L points constitute the desired result from linear convolution. This segmentation of the input data and the fitting of the output data blocks together to form the output sequence are graphically illustrated in Fig. 3.1.

498

The Discrete Fourier Transform: Its Properties and Applications

Input signal

L

L

L

x1(n) M−1 zeros

L

M−1 x2(n)

Output signal x3(n) y1(n) Discard M−1 points

y2(n) Discard M−1 points

y3(n)

Figure 3.1

Discard M−1 points

Linear FIR filtering by the overlap-save method.

Overlap-add method. In this method the size of the input data block is L points and the size of the DFTs and IDFT is N = L+M −1. To each data block we append M −1 zeros and compute the N -point DFT. Thus the data blocks may be represented as x1 (n) = {x(0), x(1), . . . , x(L − 1), 0, 0, . . . , 0}   

(3.13)

M−1 zeros

x2 (n) = {x(L), x(L + 1), . . . , x(2L − 1), 0, 0, . . . , 0}   

(3.14)

M−1 zeros

x3 (n) = {x(2L), . . . , x(3L − 1), 0, 0, . . . , 0}   

(3.15)

M−1 zeros

and so on. The two N -point DFTs are multiplied together to form Ym (k) = H (k)Xm (k),

k = 0, 1, . . . , N − 1

(3.16)

The IDFT yields data blocks of length N that are free of aliasing, since the size of the DFTs and IDFT is N = L + M − 1 and the sequences are increased to N -points by appending zeros to each block. Since each data block is terminated with M − 1 zeros, the last M − 1 points from each output block must be overlapped and added to the first M − 1 points of

499

The Discrete Fourier Transform: Its Properties and Applications

Input data L

L

L

x1(n) M−1 zeros x2(n) M−1 zeros Output data x3(n)

y1(n) M − 1 points add together

Figure 3.2

Linear FIR filtering by the overlap-add method.

y2(n) M − 1 points add together

y3(n)

the succeeding block. Hence this method is called the overlap-add method. This overlapping and adding yields the output sequence y(n) = {y1 (0), y1 (1), . . . , y1 (L − 1), y1 (L) + y2 (0), y1 (L + 1) + y2 (1), . . . , y1 (N − 1) + y2 (M − 1), y2 (M), . . .}

(3.17)

The segmentation of the input data into blocks and the fitting of the output data blocks to form the output sequence are graphically illustrated in Fig. 3.2. At this point, it may appear to the reader that the use of the DFT in linear FIR filtering not only is an indirect method of computing the output of an FIR filter, but also may be more expensive computationally, since the input data must first be converted to the frequency domain via the DFT, multiplied by the DFT of the FIR filter, and finally, converted back to the time domain via the IDFT. On the contrary, however, by using the fast Fourier transform algorithm, the DFTs and IDFT require fewer computations to compute the output sequence than the direct realization of the FIR filter in the time domain. This computational efficiency is the basic advantage of using the DFT to compute the output of an FIR filter.

4

Frequency Analysis of Signals Using the DFT To compute the spectrum of either a continuous-time or discrete-time signal, the values of the signal for all time are required. However, in practice, we observe signals for only a finite duration. Consequently, the spectrum of a signal can only be

500

The Discrete Fourier Transform: Its Properties and Applications

approximated from a finite data record. In this section we examine the implications of a finite data record in frequency analysis using the DFT. If the signal to be analyzed is an analog signal, we would first pass it through an antialiasing filter and then sample it at a rate Fs ≥ 2B , where B is the bandwidth of the filtered signal. Thus the highest frequency that is contained in the sampled signal is Fs /2. Finally, for practical purposes, we limit the duration of the signal to the time interval T0 = LT , where L is the number of samples and T is the sample interval. As we shall observe in the following discussion, the finite observation interval for the signal places a limit on the frequency resolution; that is, it limits our ability to distinguish two frequency components that are separated by less than 1/T0 = 1/LT in frequency. Let {x(n)} denote the sequence to be analyzed. Limiting the duration of the sequence to L samples, in the interval 0 ≤ n ≤ L − 1, is equivalent to multiplying {x(n)} by a rectangular window w(n) of length L. That is, x(n) ˆ = x(n)w(n) where

 w(n) =

1, 0,

0≤n≤L−1 otherwise

(4.1)

(4.2)

Now suppose that the sequence x(n) consists of a single sinusoid, that is, x(n) = cos ω0 n

(4.3)

Then the Fourier transform of the finite-duration sequence x(n) can be expressed as 1 ˆ X(ω) = [W (ω − ω0 ) + W (ω + ω0 )] 2

(4.4)

where W (ω) is the Fourier transform of the window sequence, which is (for the rectangular window) sin(ωL/2) −j ω(L−1)/2 e W (ω) = (4.5) sin(ω/2) ˆ To compute X(ω) we use the DFT. By padding the sequence x(n) ˆ with N − L zeros, we can compute the N -point DFT of the truncated (L points) sequence {x(n)}. ˆ The ˆ ˆ k )| for ωk = 2π k/N , k = 0, 1, . . . , N , is illustrated magnitude spectrum |X(k)| = |X(ω ˆ in Fig. 4.1 for L = 25 and N = 2048. We note that the windowed spectrum X(ω) is not localized to a single frequency, but instead it is spread out over the whole frequency range. Thus the power of the original signal sequence {x(n)} that was concentrated at a single frequency has been spread by the window into the entire frequency range. We say that the power has “leaked out” into the entire frequency range. Consequently, this phenomenon, which is a characteristic of windowing the signal, is called leakage.

501

The Discrete Fourier Transform: Its Properties and Applications

Figure 4.1

Magnitude spectrum for L = 25 and N = 2048, illustrating the occurrence of leakage.

Windowing not only distorts the spectral estimate due to the leakage effects, it also reduces spectral resolution. To illustrate this problem, let us consider a signal sequence consisting of two frequency components, x(n) = cos ω1 n + cos ω2 n

(4.6)

When this sequence is truncated to L samples in the range 0 ≤ n ≤ L − 1, the windowed spectrum is 1 ˆ X(ω) = [W (ω − ω1 ) + W (ω − ω2 ) + W (ω + ω1 ) + W (ω + ω2 )] 2

(4.7)

Figure 4.2 Magnitude spectrum for the signal given by (4.8), as observed through a rect angular

window .

502

The Discrete Fourier Transform: Its Properties and Applications

The spectrum W (ω) of the rectangular window sequence has its first zero crossing at ω = 2π/L. Now if |ω1 − ω2 | < 2π/L, the two window functions W (ω − ω1 ) and W (ω − ω2 ) overlap and, as a consequence, the two spectral lines in x(n) are not distinguishable. Only if (ω1 − ω2 ) ≥ 2π/L will we see two separate lobes in the ˆ spectrum X(ω). Thus our ability to resolve spectral lines of different frequencies is limited by the window main lobe width. Figure 4.2 illustrates the magnitude ˆ spectrum |X(ω)|, computed via the DFT, for the sequence x(n) = cos ω0 n + cos ω1 n + cos ω2 n

(4.8)

where ω0 = 0.2π , ω1 = 0.22π , and ω2 = 0.6π . The window lengths selected are L = 25, 50, and 100. Note that ω0 and ω1 are not resolvable for L = 25 and 50, but they are resolvable for L = 100. To reduce leakage, we can select a data window w(n) that has lower sidelobes in the frequency domain compared with the rectangular window. However, a reduction of the sidelobes in a window W (ω) is obtained at the expense of an increase in the width of the main lobe of W (ω) and hence a loss in resolution. To illustrate this point, let us consider the Hanning window, which is specified as  w(n) =

1 2 (1

0,

2π − cos L−1 n),

0≤n≤L−1 otherwise

(4.9)

ˆ Figure 4.3 shows |X(ω)| given by (4.4) for the window of (4.9). Its sidelobes are significantly smaller than those of the rectangular window, but its main lobe is approximately twice as wide. Figure 4.4 shows the spectrum of the signal in (4.8), after it is windowed by the Hanning window, for L = 50, 75, and 100. The reduction of the sidelobes and the decrease in the resolution, compared with the rectangular window, is clearly evident. For a general signal sequence {x(n)}, the frequency-domain relationship between the windowed sequence x(n) ˆ and the original sequence x(n) is given by the convolution formula  π 1 ˆ (4.10) X(ω) = X(θ)W (ω − θ)dθ 2π −π

Figure 4.3

Magnitude spectrum of the Hanning window.

503

The Discrete Fourier Transform: Its Properties and Applications

Figure 4.4 Magnitude spectrum of the signal in (4.8) as observed through a Hanning window.

The DFT of the windowed sequence x(n) ˆ is the sampled version of the spectrum ˆ X(ω). Thus we have ˆ ˆ X(k) ≡ X(ω)| ω=2πk/N    π 1 2π k = X(θ)W − θ dθ, 2π −π N

k = 0, 1, . . . , N − 1

(4.11)

Just as in the case of the sinusoidal sequence, if the spectrum of the window is relatively narrow in width compared to the spectrum X(ω) of the signal, the window function has only a small (smoothing) effect on the spectrum X(ω). On the other hand, if the window function has a wide spectrum compared to the width of X(ω), as would be the case when the number of samples L is small, the window spectrum masks the signal spectrum and, consequently, the DFT of the data reflects the spectral characteristics of the window function. Of course, this situation should be avoided. EXAMPLE 4.1 The exponential signal

 xa (t) =

e−t , 0,

t ≥0 t L where a = 0.95 and L = 10. (a) Compute and plot the signal x(n). (b) Show that X(ω) =

∞ 

x(n)e−j ωn = x(0) + 2

n=−∞

L 

x(n) cos ωn

n−1

Plot X(ω) by computing it at ω = π k/100, k = 0, 1, . . . , 100. (c) Compute   1 2π K , k = 0, 1, . . . , N − 1 ck = X N N for N = 30. (d) Determine and plot the signal x(n) ˜ =

N−1 

ckej (2π/N)kn

k=0

What is the relation between the signals x(n) and x(n)? ˜ Explain. " x(n−lN ), −L ≤ n ≤ L for N = 30. (e) Compute and plot the signal x˜ 1 (n) = ∞ l=−∞ Compare the signals x(n) ˜ and x˜ 1 (n). (f) Repeat parts (c) to (e) for N = 15.

519

The Discrete Fourier Transform: Its Properties and Applications

29

Frequency-domain sampling The signal x(n) = a |n| , −1 < a < 1 has a Fourier transform 1 − a2 X(ω) = 1 − 2a cos ω + a 2 (a) Plot X(ω) for 0 ≤ ω ≤ 2π , a = 0.8. Reconstruct and plot X(ω) from its samples X(2π k/N ), 0 ≤ k ≤ N − 1 for (b) N = 20 (c) N = 100 (d) Compare the spectra obtained in parts (b) and (c) with the original spectrum X(ω) and explain the differences. (e) Illustrate the time-domain aliasing when N = 20.

30 Frequency analysis of amplitude-modulated discrete-time signal The discrete-time signal x(n) = cos 2πf1 n + cos 2πf2 n where f1 =

1 18

and f2 =

5 128 ,

modulates the amplitude of the carrier xc (n) = cos 2πfc n

where fc =

50 128 .

The resulting amplitude-modulated signal is xam (n) = x(n) cos 2πfc n

(a) Sketch the signals x(n), xc (n), and xam (n), 0 ≤ n ≤ 255. (b) Compute and sketch the 128-point DFT of the signal xam (n), 0 ≤ n ≤ 127. (c) Compute and sketch the 128-point DFT of the signal xam (n), 0 ≤ n ≤ 99. (d) Compute and sketch the 256-point DFT of the signal xam (n), 0 ≤ n ≤ 179. (e) Explain the results obtained in parts (b) through (d), by deriving the spectrum of the amplitude-modulated signal and comparing it with the experimental results. 31

The sawtooth waveform in Fig. P31 can be expressed in the form of a Fourier series as   2 1 1 1 x(t) = sin π t − sin 2π t + sin 3π t − sin 4π t · · · π 2 3 4 (a) Determine the Fourier series coefficients ck . (b) Use an N -point subroutine to generate samples of this signal in the time domain using the first six terms of the expansion for N = 64 and N = 128. Plot the signal x(t) and the samples generated, and comment on the results.

520

The Discrete Fourier Transform: Its Properties and Applications

x(t) 1

−1

Figure P31

0

1

2

3

4

t

−1

32 Recall that the Fourier transform of x(t) = ej 0 t is X(j ) = 2π δ( − 0 ) and the Fourier transform of  1, 0 ≤ t ≤ T0 p(t) = 0, otherwise is

sin T0 /2 −j T0 /2 e T0 /2 (a) Determine the Fourier transform Y (j ) of P (j ) = T0

y(t) = p(t)ej 0 t and roughly sketch |Y (j )| versus . (b) Now consider the exponential sequence x(n) = ej ω0 n where ω0 is some arbitrary frequency in the range 0 < ω0 < π radians. Give the most general condition that ω0 must satisfy in order for x(n) to be periodic with period P (P is a positive integer). (c) Let y(n) be the finite-duration sequence y(n) = x(n)wN (n) = ej ω0 n wN (n) where wN (n) is a finite-duration rectangular sequence of length N and where x(n) is not necessarily periodic. Determine Y (ω) and roughly sketch |Y (ω)| for 0 ≤ ω ≤ 2π . What effect does N have in |Y (ω)|? Briefly comment on the similarities and differences between |Y (ω)| and |Y (j )|. (d) Suppose that x(n) = ej (2π/P )n ,

P a positive integer

and y(n) = wN (n)x(n) where N = lP , l a positive integer. Determine and sketch the N -point DFT of y(n). Relate your answer to the characteristics of |Y (ω)|. (e) Is the frequency sampling for the DFT in part (d) adequate for obtaining a rough approximation of |Y (ω)| directly from the magnitude of the DFT sequence |Y (k)|? If not, explain briefly how the sampling can be increased so that it will be possible to obtain a rough sketch of |Y (ω)| from an appropriate sequence |Y (k)|.

521

The Discrete Fourier Transform: Its Properties and Applications

33 Develop an algorithm that computes the DCT using the DFT as described in Sections 5.1 and 5.2. 34 Use the algorithm developed in Problem 33 to reproduce the results in Example 5.2. 35 Repeat Example 5.2 using the signal x(n) = a n cos(2πf0 n + φ) with a = 0.8, f0 = 0.05, and N = 32.

Answers to Selected Problems 1 2

Since x(n) is real, the real part of the DFT is even, imaginary part odd. Thus, the remaining points are {0.125 + j 0.0518, 0, 0.125 + j 0.3018}   |l|, |l| ≤ 7 (a) x˜ 2 (l) = sin 3π 8 Therefore, x1 (n) 8 x2 (n) = {1.25, 2.55, 2.55, 1.25, 0.25, −1.06, −1.06, 0.25} N2 (c) R˜ xx (k) = X1 (k)X1∗ (k)  2π = 4 [δ(k − 1) + δ(k + 1)] N ⇒ r˜xx (n) = 2 cos N n

5

N2 (d) R˜ yy (k) = X2 (k)X2∗ (k)  2π = 4 [δ(k − 1) + δ(k + 1)] N ⇒ r˜yy (n) = 2 cos N n 2 N−1 N−1  j 2π n 2π 1 ∗ N (a) + e−j N n = n=0 x1 (n)x2 (n) = 4 n=0 e

9

X3 (k) = {17, 19, 22, 19}

12

(a) s(k) = W2k X(k) s(n) = {3, 4, 0, 0, 1, 2}

14

(a) y(n) = x1 (n) 5 x2 (n) = {4, 0, 1, 2, 3}

21

(a) Fs ≡ FN = 2B = 6000 samples/sec (b) L = 120 samples × 120 = 0.02 seconds N−1 2π (a) X(k) = n=0 δ(n)e−j N kn = 1, 0 ≤ k ≤ N − 1 (c) LT =

23

1 6000

(e) X(k) = N δ(k − k0 ) (h) X(k) =

1 1−e

25 31 32

(a) X(w) = 3 + 2 cos(2w) + 4 cos(4w)   2 1 (a) ck = π2 , − π1 , 3π , − 2π ···   T0 ( − 0 ) 0) e−j 2 (a) Y (j ) = T0 sin c T0 ( − 2 (c) Y (w) =

522

−j 2π N k

N−1 sin N 2 (w−w0 ) −j 2 (w−w0 ) e w−w sin 2 0

N 2

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

The discrete fourier transform (DFT) plays an important role in many applications of digital signal processing, including linear filtering, correlation analysis, and spectrum analysis. A major reason for its importance is the existence of efficient algorithms for computing the DFT. The main topic of this chapter is the description of computationally efficient algorithms for evaluating the DFT. Two different approaches are described. One is a divide-and-conquer approach in which a DFT of size N , where N is a composite number, is reduced to the computation of smaller DFTs from which the larger DFT is computed. In particular, we present important computational algorithms, called fast Fourier transform (FFT) algorithms, for computing the DFT when the size N is a power of 2 and when it is a power of 4. The second approach is based on the formulation of the DFT as a linear filtering operation on the data. This approach leads to two algorithms, the Goertzel algorithm and the chirp-z transform algorithm, for computing the DFT via linear filtering of the data sequence.

1

Efficient Computation of the DFT: FFT Algorithms In this section we present several methods for computing the DFT efficiently. In view of the importance of the DFT in various digital signal processing applications, such as linear filtering, correlation analysis, and spectrum analysis, its efficient computation is a topic that has received considerable attention by many mathematicians, engineers, and applied scientists. Basically, the computational problem for the DFT is to compute the sequence {X(k)} of N complex-valued numbers given another sequence of data {x(n)} of length

From Chapter 8 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

523

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

N , according to the formula X(k) =

N−1 

x(n)WNkn ,

0≤k ≤N −1

(1.1)

n=0

where

WN = e−j 2π/N

(1.2)

In general, the data sequence x(n) is also assumed to be complex valued. Similarly, the IDFT becomes x(n) =

N−1 1  X(k)WN−nk , N

0≤n≤N −1

(1.3)

k=0

Since the DFT and IDFT involve basically the same type of computations, our discussion of efficient computational algorithms for the DFT applies as well to the efficient computation of the IDFT. We observe that for each value of k, direct computation of X(k) involves N complex multiplications (4N real multiplications) and N −1 complex additions (4N − 2 real additions). Consequently, to compute all N values of the DFT requires N 2 complex multiplications and N 2 − N complex additions. Direct computation of the DFT is basically inefficient, primarily because it does not exploit the symmetry and periodicity properties of the phase factor WN . In particular, these two properties are: k+N/2

= −WNk

Symmetry property:

WN

Periodicity property:

WNk+N = WNk

(1.4) ( 1.5)

The computationally efficient algorithms described in this section, known collectively as fast Fourier transform (FFT) algorithms, exploit these two basic properties of the phase factor.

1.1

Direct Computation of the DFT

For a complex-valued sequence x(n) of N points, the DFT may be expressed as  N−1  2π kn 2π kn xR (n) cos (1.6) + xI (n) sin XR (k) = N N n=0

XI (k) = −

N−1  n=0

2π kn 2π kn xR (n) sin − xI (n) cos N N

The direct computation of (1.6) and (1.7) requires: 1. 2N 2 evaluations of trigonometric functions. 2. 4N 2 real multiplications. 3. 4N (N − 1) real additions. 4. A number of indexing and addressing operations.

524

 ( 1.7)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

These operations are typical of DFT computational algorithms. The operations in items 2 and 3 result in the DFT values XR (k) and XI (k). The indexing and addressing operations are necessary to fetch the data x(n), 0 ≤ n ≤ N − 1, and the phase factors and to store the results. The variety of DFT algorithms optimize each of these computational processes in a different way.

1.2

Divide-and-Conquer Approach to Computation of the DFT

The development of computationally efficient algorithms for the DFT is made possible if we adopt a divide-and-conquer approach. This approach is based on the decomposition of an N -point DFT into successively smaller DFTs. This basic approach leads to a family of computationally efficient algorithms known collectively as FFT algorithms. To illustrate the basic notions, let us consider the computation of an N -point DFT, where N can be factored as a product of two integers, that is, N = LM

(1.8)

The assumption that N is not a prime number is not restrictive, since we can pad any sequence with zeros to ensure a factorization of the form (1.8). Now the sequence x(n), 0 ≤ n ≤ N −1, can be stored either in a one-dimensional array indexed by n or as a two-dimensional array indexed by l and m, where 0 ≤ l ≤ L − 1 and 0 ≤ m ≤ M − 1 as illustrated in Fig. 1.1. Note that l is the row index and m is the column index. Thus, the sequence x(n) can be stored in a rectangular n

0

1

x(0)

x(1)

x(2)



N−1



x(N − 1)

(a)

l

column index M−1

0

x(0, 0)

x(0, 1)



1

x(1, 0)

x(1, 1)



2

x(2, 0)

x(2, 1)









1



0

L−1



m row index

… (b)

Figure 1.1 Two dimensional data array for storing the sequence

x(n), 0 ≤ n ≤ N − 1.

525

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

array in a variety of ways, each of which depends on the mapping of index n to the indexes (l, m). For example, suppose that we select the mapping n = Ml + m

(1.9)

This leads to an arrangement in which the first row consists of the first M elements of x(n), the second row consists of the next M elements of x(n), and so on, as illustrated in Fig. 1.2(a). On the other hand, the mapping n = l + mL

(1.10)

stores the first L elements of x(n) in the first column, the next L elements in the second column, and so on, as illustrated in Fig. 1.2(b). n = Ml + m

Row-wise m

0

x(0)

x(1)

x(2)



x(M − 1)

1

x(M)

x(M + 1)

x(M + 2)



x(2M − 1)

2

x(2M)

x(2M + 1)

x(2M + 2)



x(3M − 1)





x(LM − 1)

L−1

x((L − 1)M)



2



1



M−1

0



l

x((L −1)M +1) x((L −1)M +2) (a)

n = l + mL

Column-wise m l

x(L)

x(2L)



x((M − 1)L)

1

x(1)

x(L + 1)

x(2L + 1)



x((M − 1)L +1)

2

x(2)

x(L + 2)

x(2L + 2)



x((M − 1)L + 2)



x(L − 1)

x(2L − 1)

x(3L − 1)



(b)

Figure 1.2 Two arrangements for the data arrays.



x(0)



0



2



1

L−1

526

M−1

0

x(LM − 1)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

A similar arrangement can be used to store the computed DFT values. In particular, the mapping is from the index k to a pair of indices (p, q), where 0 ≤ p ≤ L − 1 and 0 ≤ q ≤ M − 1. If we select the mapping k = Mp + q

(1.11)

the DFT is stored on a row-wise basis, where the first row contains the first M elements of the DFT X(k), the second row contains the next set of M elements, and so on. On the other hand, the mapping k = qL + p

(1.12)

results in a column-wise storage of X(k), where the first L elements are stored in the first column, the second set of L elements are stored in the second column, and so on. Now suppose that x(n) is mapped into the rectangular array x(l, m) and X(k) is mapped into a corresponding rectangular array X(p, q). Then the DFT can be expressed as a double sum over the elements of the rectangular array multiplied by the corresponding phase factors. To be specific, let us adopt a column-wise mapping for x(n) given by (1.10) and the row-wise mapping for the DFT given by (1.11). Then M−1  L−1  (Mp+q)(mL+l) X(p, q) = (1.13) x(l, m)WN m=0 l=0

But (Mp+q)(mL+l)

WN Nmp

mqL

MLmp

= WN

mq

mq

mLq

WN

Mpl

lq

(1.14)

WN WN Mpl

pl

pl

However, WN = 1, WN = WN/L = WM , and WN = WN/M = WL . With these simplications, (1.13) can be expressed as X(p, q) =

L−1 



l=0

lq WN

M−1 

 mq x(l, m)WM

lp

WL

(1.15)

m=0

The expression in (1.15) involves the computation of DFTs of length M and length L. To elaborate, let us subdivide the computation into three steps: 1. First, we compute the M -point DFTs F (l, q) ≡

M−1 

mq

x(l, m)WM ,

0≤q ≤M −1

(1.16)

m=0

for each of the rows l = 0, 1, . . . , L − 1. 2. Second, we compute a new rectangular array G(l, q) defined as lq

G(l, q) = WN F (l, q),

0≤l ≤L−1 0≤q ≤M −1

(1.17)

527

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

3. Finally, we compute the L-point DFTs X(p, q) =

L−1 

lp

G(l, q)WL

(1.18)

l=0

for each column q = 0, 1, . . . , M − 1, of the array G(l, q). On the surface it may appear that the computational procedure outlined above is more complex than the direct computation of the DFT. However, let us evaluate the computational complexity of (1.15). The first step involves the computation of L DFTs, each of M points. Hence this step requires LM 2 complex multiplications and LM(M − 1) complex additions. The second step requires LM complex multiplications. Finally, the third step in the computation requires ML2 complex multiplications and ML(L − 1) complex additions. Therefore, the computational complexity is Complex multiplications: N (M + L + 1) (1.19) Complex additions: N (M + L − 2) where N = ML. Thus the number of multiplications has been reduced from N 2 to N (M + L + 1) and the number of additions has been reduced from N (N − 1) to N (M + L − 2). For example, suppose that N = 1000 and we select L = 2 and M = 500. Then, instead of having to perform 106 complex multiplications via direct computation of the DFT, this approach leads to 503,000 complex multiplications. This represents a reduction by approximately a factor of 2. The number of additions is also reduced by about a factor of 2. When N is a highly composite number, that is, N can be factored into a product of prime numbers of the form (1.20) N = r1 r2 · · · rν then the decomposition above can be repeated (ν − 1) more times. This procedure results in smaller DFTs, which, in turn, leads to a more efficient computational algorithm. In effect, the first segmentation of the sequence x(n) into a rectangular array of M columns with L elements in each column resulted in DFTs of sizes L and M . Further decomposition of the data in effect involves the segmentation of each row (or column) into smaller rectangular arrays which result in smaller DFTs. This procedure terminates when N is factored into its prime factors. EXAMPLE 1.1 To illustrate this computational procedure, let us consider the computation of an N = 15 point DFT. Since N = 5 × 3 = 15, we select L = 5 and M = 3. In other words, we store the 15-point sequence x(n) column-wise as follows: Row 1: x(0, 0) = x(0) x(0, 1) = x(5) x(0, 2) = x(10) Row 2: x(1, 0) = x(1) x(1, 1) = x(6) x(1, 2) = x(11) Row 3: x(2, 0) = x(2) x(2, 1) = x(7) x(2, 2) = x(12) Row 4: x(3, 0) = x(3) x(3, 1) = x(8) x(3, 2) = x(13) Row 5: x(4, 0) = x(4) x(4, 1) = x(9) x(4, 2) = x(14)

528

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Now, we compute the three-point DFTs for each of the five rows. This leads to the following 5 × 3 array: F (0, 0) F (0, 1) F (0, 2) F (1, 0) F (1, 1) F (1, 2) F (2, 0) F (2, 1) F (2, 2) F (3, 0) F (3, 1) F (3, 2) F (4, 0) F (4, 1) F (4, 2) lq

lq

The next step is to multiply each of the terms F (l, q) by the phase factors WN = W15 , 0 ≤ l ≤ 4 and 0 ≤ q ≤ 2. This computation results in the 5 × 3 array: Column 1 G(0, 0) G(1, 0) G(2, 0) G(3, 0) G(4, 0)

Column 2 G(0, 1) G(1, 1) G(2, 1) G(3, 1) G(4, 1)

Column 3 G(0, 2) G(1, 2) G(2, 2) G(3, 2) G(4, 2)

The final step is to compute the five-point DFTs for each of the three columns. This computation yields the desired values of the DFT in the form X(0, 0) = X(0) X(1, 0) = X(3) X(2, 0) = X(6) X(3, 0) = X(9) X(4, 0) = X(12)

X(0, 1) = X(1) X(1, 1) = X(4) X(2, 1) = X(7) X(3, 1) = X(10) X(4, 1) = X(13)

X(0, 2) = X(2) X(1, 2) = X(5) X(2, 2) = X(8) X(3, 2) = X(11) X(4, 2) = X(14)

Figure 1.3 illustrates the steps in the computation. It is interesting to view the segmented data sequence and the resulting DFT in terms of one-dimensional arrays. When the input sequence x(n) and the output DFT X(k) in the two-dimensional arrays are read across from row 1 through row 5, we obtain the following sequences: INPUT ARRAY x(0) x(5) x(10) x(1) x(6) x(11) x(2) x(7) x(12) x(3) x(8) x(13) x(4) x(9) x(14) OUTPUT ARRAY X(0) X(1) X(2) X(3) X(4) X(5) X(6) X(7) X(8) X(9) X(10) X(11) X(12) X(13) X(14) lq

W 15 x(10)

0

11

x(0)

12

x(1)

13

x(2)

14

x(3)

6

1 7 2 8 3 9 4

X(2)

T

x(5)

t oin

DF

X(1) X(5)

3-P

X(0) 5-Point DFT

10 5

X(8) X(3) X(11) X(6) X(14) X(9)

x(4) X(12)

Figure 1.3 Computation of N = 15-point DFT by means of 3-point and 5-point DFTs.

529

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

We observe that the input data sequence is shuffled from the normal order in the computation of the DFT. On the other hand, the output sequence occurs in normal order. In this case the rearrangement of the input data array is due to the segmentation of the one-dimensional array into a rectangular array and the order in which the DFTs are computed. This shuffling of either the input data sequence or the output DFT sequence is a characteristic of most FFT algorithms.

To summarize, the algorithm that we have introduced involves the following computations: Algorithm 1 1. Store the signal column-wise. 2. Compute the M -point DFT of each row. lq

3. Multiply the resulting array by the phase factors WN . 4. Compute the L-point DFT of each column 5. Read the resulting array row-wise. An additional algorithm with a similar computational structure can be obtained if the input signal is stored row-wise and the resulting transformation is column-wise. In this case we select n = Ml + m (1.21) k = qL + p This choice of indices leads to the formula for the DFT in the form X(p, q) =

M−1  L−1 

pm

pl

qm

x(l, m)WN WL WM

m=0 l=0

=

M−1 

L−1  mq

WM

m=0

lp

x(l, m)WL

l=0

Thus we obtain a second algorithm. Algorithm 2 1. Store the signal row-wise. 2. Compute the L-point DFT at each column. pm

3. Multiply the resulting array by the factors WN . 4. Compute the M -point DFT of each row. 5. Read the resulting array column-wise.

530

(1.22)

 mp

WN

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

The two algorithms given above have the same complexity. However, they differ in the arrangement of the computations. In the following sections we exploit the divide-and-conquer approach to derive fast algorithms when the size of the DFT is restricted to be a power of 2 or a power of 4.

1.3

Radix-2 FFT Algorithms

In the preceding section we described four algorithms for efficient computation of the DFT based on the divide-and-conquer approach. Such an approach is applicable when the number N of data points is not a prime. In particular, the approach is very efficient when N is highly composite, that is, when N can be factored as N = r1 r2 r3 · · · rν , where the {rj } are prime. Of particular importance is the case in which r1 = r2 = · · · = rν ≡ r , so that N = r ν . In such a case the DFTs are of size r , so that the computation of the N -point DFT has a regular pattern. The number r is called the radix of the FFT algorithm. In this section we describe radix-2 algorithms, which are by far the most widely used FFT algorithms. Radix-4 algorithms are described in the following section. Let us consider the computation of the N = 2ν point DFT by the divide-andconquer approach specified by (1.16) through (1.18). We select M = N/2 and L = 2. This selection results in a split of the N -point data sequence into two N/2point data sequences f1 (n) and f2 (n), corresponding to the even-numbered and odd-numbered samples of x(n), respectively, that is, f1 (n) = x(2n) (1.23) N −1 2 Thus f1 (n) and f2 (n) are obtained by decimating x(n) by a factor of 2, and hence the resulting FFT algorithm is called a decimation-in-time algorithm. Now the N -point DFT can be expressed in terms of the DFTs of the decimated sequences as follows: f2 (n) = x(2n + 1),

X(k) =

N−1 

x(n)WNkn ,

n = 0, 1, . . . ,

k = 0, 1, . . . , N − 1

n=0

=



x(n)WNkn +

n even





(1.24)

n odd

(N/2)−1

=

x(n)WNkn 

(N/2)−1

x(2m)WN2mk +

m=0

x(2m + 1)WNk(2m+1)

m=0

But WN2 = WN/2 . With this substitution, (1.24) can be expressed as 



(N/2)−1

(N/2)−1

X(k) =

km f1 (m)WN/2 + WNk

m=0

= F1 (k) + WNk F2 (k),

km f2 (m)WN/2

m=0

(1.25)

k = 0, 1, . . . , N − 1

531

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

where F1 (k) and F2 (k) are the N/2-point DFTs of the sequences f1 (m) and f2 (m), respectively. Since F1 (k) and F2 (k) are periodic, with period N/2, we have F1 (k + N/2) = k+N/2 F1 (k) and F2 (k + N/2) = F2 (k). In addition, the factor WN = −WNk . Hence (1.25) can be expressed as X(k) = F1 (k) + WNk F2 (k),

k = 0, 1, . . . ,

(1.26)

 N X k+ = F1 (k) − WNk F2 (k), 2

N −1 2

k = 0, 1, . . . ,

N −1 2

(1.27)

We observe that the direct computation of F1 (k) requires (N/2)2 complex multiplications. The same applies to the computation of F2 (k). Furthermore, there are N/2 additional complex multiplications required to compute WNk F2 (k). Hence the computation of X(k) requires 2(N/2)2 + N/2 = N 2 /2 + N/2 complex multiplications. This first step results in a reduction of the number of multiplications from N 2 to N 2 /2 + N/2, which is about a factor of 2 for N large. To be consistent with our previous notation, we may define ,

k = 0, 1, . . . ,

N −1 2

G2 (k) = WNk F2 (k),

k = 0, 1, . . . ,

N −1 2

G1 (k) = F1 (k)

Then the DFT X(k) may be expressed as X(k) = G1 (k) + G2 (k), N X(k + ) = G1 (k) − G2 (k), 2

k = 0, 1, . . . ,

N −1 2

N k = 0, 1, . . . , − 1 2

(1.28)

This computation is illustrated in Fig. 1.4. Having performed the decimation-in-time once, we can repeat the process for each of the sequences f1 (n) and f2 (n). Thus f1 (n) would result in the two N /4-point sequences N n = 0, 1, . . . , − 1 v11 (n) = f1 (2n), 4 (1.29) N v12 (n) = f1 (2n + 1), n = 0, 1, . . . , − 1 4 and f2 (n) would yield v21 (n) = f2 (2n), v22 (n) = f2 (2n + 1),

532

N −1 4 N n = 0, 1, . . . , − 1 4

n = 0, 1, . . . ,

(1.30)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

x(N − 2)

x(0) x(2) x(4) x(1) x(3)

N/2-Point DFT

( )

F1 N − 1 2

F1(0) F1(1) F1(2)

G1(k)

F2(0) F2(1) Phase factors

( )

X N −1 2

2-Point DFT

k WN

G2(k)

X(0) X(1) X(N −1)

() ( )

X N 2

X N +1 2

Figure 1.4 First step in the decimation-in-time algorithm.

By computing N /4-point DFTs, we would obtain the N /2-point DFTs F1 (k) and F2 (k) from the relations k = 0, 1, . . . ,

N −1 4

k = V11 (k) − WN/2 V12 (k),

k = 0, 1, . . . ,

N −1 4

k F2 (k) = V21 (k) + WN/2 V22 (k),

k = 0, 1, . . . ,

N −1 4

k V12 (k), F1 (k) = V11 (k) + WN/2

 F1

 F2

N k+ 4

N k+ 4



k V22 (k), = V21 (k) − WN/2

k = 0, . . . ,

N −1 4

(1.31)

(1.32)

where the {Vij (k)} are the N /4-point DFTs of the sequences {vij (n)}. We observe that the computation of {Vij (k)} requires 4(N/4)2 multiplications and hence the computation of F1 (k) and F2 (k) can be accomplished with N 2 /4+N/2 complex multiplications. An additional N/2 complex multiplications are required to compute X(k) from F1 (k) and F2 (k). Consequently, the total number of multiplications is reduced approximately by a factor of 2 again to N 2 /4 + N . The decimation of the data sequence can be repeated again and again until the resulting sequences are reduced to one-point sequences. For N = 2ν , this decimation can be performed ν = log2 N times. Thus the total number of complex multiplications is reduced to (N/2) log2 N . The number of complex additions is N log2 N . Table 1 presents a comparison of the number of complex multiplications in the FFT and in the direct computation of the DFT. For illustrative purposes, Fig. 1.5 depicts the computation of an N = 8-point DFT. We observe that the computation is performed in three stages, beginning with the computations of four two-point DFTs, then two four-point DFTs, and finally, one

533

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

TABLE 1 Comparison of Computational Complexity for the Direct Computation of the

DFT Versus the FFT Algorithm Number of Complex Multiplications Points, in Direct Computation, N N2 4 16 8 64 16 256 32 1,024 64 4,096 128 16,384 256 65,536 512 262,144 1,024 1,048,576

Complex Multiplications in FFT Algorithm, (N/2) log2 N 4 12 32 80 192 448 1,024 2,304 5,120

Speed Improvement Factor 4.0 5.3 8.0 12.8 21.3 36.6 64.0 113.8 204.8

eight-point DFT. The combination of the smaller DFTs to form the larger DFT is illustrated in Fig. 1.6 for N = 8. Observe that the basic computation performed at every stage, as illustrated in Fig. 1.6, is to take two complex numbers, say the pair (a, b), multiply b by WNr , and then add and subtract the product from a to form two new complex numbers (A, B). This basic computation, which is shown in Fig. 1.7, is called a butterfly because the flow graph resembles a butterfly. In general, each butterfly involves one complex multiplication and two complex additions. For N = 2ν , there are N/2 butterflies per stage of the computation process and log2 N stages. Therefore, as previously indicated, the total number of complex multiplications is (N/2) log2 N and complex additions is N log2 N . Once a butterfly operation is performed on a pair of complex numbers (a, b) to produce (A, B), there is no need to save the input pair (a, b). Hence we can store the result (A, B) in the same locations as (a, b). Consequently, we require a fixed amount of storage, namely, 2N storage registers, in order to store the results x(0) x(4)

2-point DFT

x(2) x(6)

2-point DFT

x(1) x(5)

2-point DFT

x(3) x(7)

2-point DFT

Combine 2-point DFT’s

X(0) X(1) X(2) Combine 4-point DFT’s

Combine 2-point DFT’s

X(3) X(4) X(5) X(6) X(7)

Figure 1.5 Three stages in the computation of an N = 8-point DFT.

534

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Stage 1

Stage 2

Stage 3

x(0)

X(0) 0

x(4)

W8

W 08

x(2)

x(6)

X(1)

−1

W 08

W 28

X(3)

−1

−1

W 08

x(1)

−1 0

x(5)

X(4)

1

W8

W8 −1

−1 W 08

x(3)

W 28 −1

0

x(7)

X(2)

−1

2

W8

−1 W 38

W8 −1

−1

−1

X(5)

X(6)

X(7)

Figure 1.6 Eight-point decimation-in-time FFT algorithm.

(N complex numbers) of the computations at each stage. Since the same 2N storage locations are used throughout the computation of the N -point DFT, we say that the computations are done in place. A second important observation is concerned with the order of the input data sequence after it is decimated (ν − 1) times. For example, if we consider the case where N = 8, we know that the first decimation yields the sequence x(0), x(2), x(4), x(6), x(1), x(3), x(5), x(7), and the second decimation results in the sequence x(0), x(4), x(2), x(6), x(1), x(5), x(3), x(7). This shuffling of the input data sequence has a well-defined order as can be ascertained from observing Fig. 1.8, which illustrates the decimation of the eight-point sequence. By expressing the index n, in the sequence x(n), in binary form, we note that the order of the decimated data sequence is easily obtained by reading the binary representation of the index n in reverse order. Thus the data point x(3) ≡ x(011) is placed in position m = 110 or m = 6 in the decimated array. Thus we say that the data x(n) after decimation is stored in bit-reversed order. A = a + WrN b

a

Figure 1.7

Basic butterfly computation in the decimation-in-time FFT algorithm.

b

WrN −1

B = a − WrN b

535

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Memory address (decimal) (binary) 000 0

Memory

Data decimation 1

Data decimation 2

x(0)

x(0)

x(0)

1

001

x(1)

x(2)

x(4)

2

010

x(2)

x(4)

x(2)

3

011

x(3)

x(6)

x(6)

4

100

x(4)

x(1)

x(1)

5

101

x(5)

x(3)

x(5)

6

110

x(6)

x(5)

x(3)

7

111

x(7)

x(7)

x(7)

Natural order (n2 n 1 n 0 ) (0 0 0) (0 0 1) (0 1 0) (0 1 1) (1 0 0) (1 0 1) (1 1 0) (1 1 1)

Bit-reversed order

(a) → (n 1n 0 n 2 )

→ (n 0 n 1 n 2 )

→ → → → → → → →

→ → → → → → → →

(0 0 0) (0 1 0) (1 0 0) (1 1 0) (0 0 1) (0 1 1) (1 0 1) (1 1 1) (b)

(0 0 0) (1 0 0) (0 1 0) (1 1 0) (0 0 1) (1 0 1) (0 1 1) (1 1 1)

Figure 1.8 Shuffling of the data and bit reversal.

With the input data sequence stored in bit-reversed order and the butterfly computations performed in place, the resulting DFT sequence X(k) is obtained in natural order (i.e., k = 0, 1, . . . , N − 1). On the other hand, we should indicate that it is possible to arrange the FFT algorithm such that the input is left in natural order and the resulting output DFT will occur in bit-reversed order. Furthermore, we can impose the restriction that both the input data x(n) and the output DFT X(k) be in natural order, and derive an FFT algorithm in which the computations are not done in place. Hence such an algorithm requires additional storage. Another important radix-2 FFT algorithm, called the decimation-in-frequency algorithm, is obtained by using the divide-and-conquer approach described in Section 1.2 with the choice of M = 2 and L = N/2. This choice of parameters implies a column-wise storage of the input data sequence. To derive the algorithm, we begin by splitting the DFT formula into two summations, of which one involves the sum over the first N /2 data points and the second the sum over the last N /2 data points.

536

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Thus we obtain 

(N/2)−1

X(k) =

x(n)WNkn

+

n=0



N−1 

x(n)WNkn

n=N/2

=



(N/2)−1

(N/2)−1 Nk/2

x(n)WNkn + WN

n=0 kN/2

Since WN

 x n+

n=0

N 2

(1.33)

WNkn

= (−1)k , the expression (1.33) can be rewritten as (N/2)−1 

X(k) =

 n=0

  N x(n) + (−1)k x n + WNkn 2

(1.34)

Now, let us split (decimate) X(k) into the even- and odd-numbered samples. Thus we obtain (N/2)−1 

X(2k) =

 n=0



N x(n) + x n + 2

 kn , WN/2

k = 0, 1, . . . ,

N −1 2

(1.35)

and (N/2)−1 

X(2k + 1) =

 n=0

  N kn x(n) − x n + , WNn WN/2 2

k = 0, 1, . . . ,

N −1 2 (1.36)

where we have used the fact that WN2 = WN/2 . If we define the N /2-point sequences g1 (n) and g2 (n) as  N g1 (n) = x(n) + x n + 2    N N g2 (n) = x(n) − x n + WNn , n = 0, 1, 2, . . . , − 1 2 2

(1.37)

then 

(N/2)−1

X(2k) =

kn g1 (n)WN/2

n=0



(1.38)

(N/2)−1

X(2k + 1) =

kn g2 (n)WN/2

n=0

The computation of the sequences g1 (n) and g2 (n) according to (1.37) and the subsequent use of these sequences to compute the N /2-point DFTs are depicted in Fig. 1.9. We observe that the basic computation in this figure involves the butterfly operation illustrated in Fig. 1.10.

537

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

X(0)

x(0)

x(1)

X(2) 4-point DFT

x(2)

X(4)

x(3)

X(6)

x(4)

x(5)

x(6)

Figure 1.9

First stage of the decimation-in-frequency FFT algorithm.

W 08

X(1)

−1

W 18

X(3)

−1

4-point DFT

W 28

X(5)

−1 3

x(7)

W8

X(7)

−1

This computational procedure can be repeated through decimation of the N /2point DFTs, X(2k) and X(2k + 1). The entire process involves ν = log2 N stages of decimation, where each stage involves N /2 butterflies of the type shown in Fig. 1.10. Consequently, the computation of the N -point DFT via the decimation-in-frequency FFT algorithm requires (N/2) log2 N complex multiplications and N log2 N complex additions, just as in the decimation-in-time algorithm. For illustrative purposes, the eight-point decimation-in-frequency algorithm is given in Fig. 1.11. We observe from Fig. 1.11 that the input data x(n) occurs in natural order, but the output DFT occurs in bit-reversed order. We also note that the computations are performed in place. However, it is possible to reconfigure the decimation-infrequency algorithm so that the input sequence occurs in bit-reversed order while the output DFT occurs in normal order. Furthermore, if we abandon the requirement that the computations be done in place, it is also possible to have both the input data and the output DFT in normal order.

Figure 1.10

Basic butterfly computation in the decimation-in-frequency FFT algorithm.

538

A=a + b

a r

b

WN −1

r

B = (a − 2b)WN

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

X(0)

x(0) W 08

x(1)

−1

W 08

x(2)

X(2)

−1

W 28

x(3)

W 08

−1

−1

W 08

x(4)

W 18

W 08

−1

−1

W 28

x(6)

W 08

3

W 28

W8

W 08

−1

−1

X(5)

X(3)

−1

−1

x(7)

X(6)

X(1)

−1

x(5)

X(4)

−1

X(7)

Figure 1.11 N = 8-point decimation-in-frequency FFT algorithm.

1.4 Radix-4 FFT Algorithms When the number of data points N in the DFT is a power of 4 (i.e., N = 4ν ), we can, of course, always use a radix-2 algorithm for the computation. However, for this case, it is more efficient computationally to employ a radix-4 FFT algorithm. Let us begin by describing a radix-4 decimation-in-time FFT algorithm, which is obtained by selecting L = 4 and M = N/4 in the divide-and-conquer approach described in Section 1.2. For this choice of L and M , we have l , p = 0, 1, 2, 3; m, q = 0, 1, . . . , N/4 − 1; n = 4m + l ; and k = (N/4)p + q . Thus we split or decimate the N -point input sequence into four subsequences, x(4n), x(4n + 1), x(4n + 2), x(4n + 3), n = 0, 1, . . . , N/4 − 1. By applying (1.15) we obtain X(p, q) =

3 

lq lp WN F (l, q) W4 ,

p = 0, 1, 2, 3

(1.39)

l=0

where F (l, q) is given by (1.16), that is, 

(N/4)−1

F (l, q) =

m=0

mq

x(l, m)WN/4 ,

l = 0, 1, 2, 3, q = 0, 1, 2, . . . ,

N −1 4

(1.40)

539

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

and x(l, m) = x(4m + l)  N p+q X(p, q) = X 4

(1.41) (1.42)

Thus, the four N /4-point DFTs obtained from (1.40) are combined according to (1.39) to yield the N -point DFT. The expression in (1.39) for combining the N /4point DFTs defines a radix-4 decimation-in-time butterfly, which can be expressed in matrix form as      W 0 F (0, q)  X(0, q) 1 1 1 1 N q WN F (1, q)  j   X(1, q)   1 −j −1   (1.43) 2q  =  X(2, q) 1 −1 1 −1  WN F (2, q)  3q X(3, q) 1 j −1 −j WN F (3, q) The radix-4 butterfly is depicted in Fig. 1.12(a) and in a more compact form in Fig. 1.12(b). Note that since WN0 = 1, each butterfly involves three complex multiplications, and 12 complex additions. This decimation-in-time procedure can be repeated recursively ν times. Hence the resulting FFT algorithm consists of ν stages, where each stage contains N /4 butterflies. Consequently, the computational burden for the algorithm is 3νN/4 = (3N/8) log2 N complex multiplications and (3N/2) log2 N complex additions. We note that the number of multiplications is reduced by 25%, but the number of additions has increased by 50% from N log2 N to (3N/2) log2 N . It is interesting to note, however, that by performing the additions in two steps, it is possible to reduce the number of additions per butterfly from 12 to 8. This can 0

WN

q

−j −1

WN

j −1

W 2q N

1 −1

0 q 2q 3q

j W 3q N

−1 −j

(a)

(b)

Figure 1.12 Basic butterfly computation in a radix-4 FFT algorithm.

540

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

be accomplished by expressing the matrix of the linear transformation in (1.43) as a product of two matrices as follows:       W 0 F (0, q)  X(0, q) 1 0 1 0 1 0 1 0 N q WN F (1, q)  0 −j   1 0 −1 0  X(1, q)   0 1   (1.44) 2q  =   X(2, q) 1 0 −1 0 0 1 0 1  WN F (2, q)  3q X(3, q) 0 1 0 j 0 1 0 −1 W F (3, q) N

Now each matrix multiplication involves four additions for a total of eight additions. Thus the total number of complex additions is reduced to N log2 N , which is identical to the radix-2 FFT algorithm. The computational savings results from the 25% reduction in the number of complex multiplications. An illustration of a radix-4 decimation-in-time FFT algorithm is shown in Fig. 1.13 for N = 16. Note that in this algorithm, the input sequence is in normal order while the output DFT is shuffled. In the radix-4 FFT algorithm, where the decimation is by a factor of 4, the order of the decimated sequence can be determined by reversing the order of the number that represents the index n in a quaternary number system (i.e., the number system based on the digits 0, 1, 2, 3). A radix-4 decimation-in-frequency FFT algorithm can be obtained by selecting L = N/4, M = 4; l , p = 0, 1, . . . , N/4 − 1; m, q = 0, 1, 2, 3; n = (N/4)m + l ; and x(0)

X(0) 0 0

x(1)

0 X(4)

x(2)

X(8) 0

x(3)

X(12)

x(4)

X(1) 0

1 2

x(5) x(6) x(7)

0 0

0 X(5) X(9) 0

0

X(13)

0

x(8)

X(2)

0 x(9) 0 x(10)

2 4

0

0 X(6) X(10) 6 X(14)

x(11) x(12) x(13) x(14) x(15)

0 0 0 0

X(3) 3 6

0 X(7) X(11) 9 X(15)

Figure 1.13 Sixteen-point radix-4 decimation-in-time algorithm with input in normal order and output in digit-reversed order. The integer multipliers shown on the graph represent the exponents on W16 .

541

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

k = 4p + q . With this choice of parameters, the general equation given by (1.15) can be expressed as 

(N/4)−1

X(p, q) =

lp

(1.45)

G(l, q)WN/4

l=0

where q = 0, 1, 2, 3

lq

G(l, q) = WN F (l, q),

l = 0, 1, . . . ,

(1.46)

N −1 4

and F (l, q) =

3 

q = 0, 1, 2, 3

mq

x(l, m)W4 ,

m=0

(1.47)

N −1 4

l = 0, 1, 2, 3, . . . ,

We note that X(p, q) = X(4p + q), q = 0, 1, 2, 3. Consequently, the N point DFT is decimated into four N/4-point DFTs and hence we have a decimationin-frequency FFT algorithm. The computations in (1.46) and (1.47) define the basic radix-4 butterfly for the decimation-in-frequency algorithm. Note that the lq multiplications by the factors WN occur after the combination of the data points x(l, m), just as in the case of the radix-2 decimation-in-frequency algorithm. A 16-point radix-4 decimation-in-frequency FFT algorithm is shown in Fig. 1.14. Its input is in normal order and its output is in digit-reversed order. It has exactly the same computational complexity as the decimation-in-time radix-4 FFT algorithm. For illustrative purposes, let us rederive the radix-4 decimation-in-frequency algorithm by breaking the N -point DFT formula into four smaller DFTs. We have

X(k) =

N−1 

x(n)WNkn

n=0



N/4−1

=

n=0





x(n)WNkn +

n=N/4

N/4−1

=



N/4−1 Nk/4

n=0

n=0



N/4−1

+

n=0



x(n)WNkn +

n=N/2

x(n)WNkn + WN

kN/2 WN



3N/4−1

N/2−1

x(n)WNkn +

N x n+ 2

N−1 

x(n)WNkn

n=3N/4

 N x n+ WNkn 4





N/4−1

WNnk

+

3kN/4 WN

n=0



3N x n+ 4

WNkn (1.48)

From the definition of the phase factors, we have kN/4

WN

542

= (−j )k ,

Nk/2

WN

= (−1)k ,

3N k/4

WN

= (j )k

(1.49)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

x(0)

X(0) 0

0

x(1) 0

0 0

x(2)

X(4) X(8)

0

0 x(3)

X(12) 0

x(4)

0

X(1) 0

x(5)

1

x(6)

2

0 0

X(5) X(9)

0 x(7)

X(13)

3 0

x(8)

X(2)

2

0

x(9) 4 x(10)

0 X(14)

0 3

x(12)

X(3) 0

6

x(13)

9

x(14)

X(6) X(10)

6

x(11)

0 0

0 0

X(7) X(11)

0 x(15)

X(15)

Figure 1.14 Sixteen-point, radix-4 decimation-in-frequency algorithm with input

in normal order and output in digit-reversed order.

After substitution of (1.49) into (1.48), we obtain N/4−1 

X(k) =

 n=0



N x(n) + (−j ) x n + 4



k



N + (−1)k x n + 4





3N + (j )k x n + 4

(1.50)

 WNnk

The relation in (1.50) is not an N/4-point DFT because the phase factor depends on N and not on N/4. To convert it into an N/4-point DFT, we subdivide the DFT sequence into four N/4-point subsequences, X(4k), X(4k + 1), X(4k + 2), and X(4k+3), k = 0, 1, . . . , N/4−1. Thus we obtain the radix-4 decimation-in-frequency DFT as N/4−1 

X(4k) =

 n=0

 N x(n) + x n + 4

(1.51)

   3N N kn +x n+ WN0 WN/4 +x n+ 2 4

543

Efficient Computation of the DFT: Fast Fourier Transform Algorithms N/4−1 

X(4k + 1) =

 n=0

 N x(n) − j x n + 4

(1.52)

   3N N kn + jx n + WNn WN/4 −x n+ 2 4 N/4−1 

X(4k + 2) =

 n=0



N x(n) − x n + 4

(1.53)

   3N N kn −x n+ WN2n WN/4 +x n+ 2 4 N/4−1 

X(4k + 3) =

 n=0



N x(n) + j x n + 4

(1.54)

   3N N kn − jx n + WN3n WN/4 −x n+ 2 4 kn where we have used the property WN4kn = WN/4 . Note that the input to each N/4point DFT is a linear combination of four signal samples scaled by a phase factor. This procedure is repeated ν times, where ν = log4 N .

1.5

Split-Radix FFT Algorithms

An inspection of the radix-2 decimation-in-frequency flowgraph shown in Fig. 1.11 indicates that the even-numbered points of the DFT can be computed independently of the odd-numbered points. This suggests the possibility of using different computational methods for independent parts of the algorithm with the objective of reducing the number of computations. The split-radix FFT (SRFFT) algorithms exploit this idea by using both a radix-2 and a radix-4 decomposition in the same FFT algorithm. We illustrate this approach with a decimation-in-frequency SRFFT algorithm due to Duhamel (1986). First, we recall that in the radix-2 decimation-in-frequency FFT algorithm, the even-numbered samples of the N -point DFT are given as N/2−1 

X(2k) =

 n=0

  N nk x(n) + x n + , WN/2 2

k = 0, 1, . . . ,

N −1 2

(1.55)

Note that these DFT points can be obtained from an N/2-point DFT without any additional multiplications. Consequently, a radix-2 suffices for this computation. The odd-numbered samples {X(2k + 1)} of the DFT require the premultiplication of the input sequence with the phase factors WNn . For these samples a radix-4 decomposition produces some computational efficiency because the four-point DFT has the largest multiplication-free butterfly. Indeed, it can be shown that using a radix greater than 4 does not result in a significant reduction in computational complexity.

544

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

If we use a radix-4 decimation-in-frequency FFT algorithm for the odd-numbered samples of the N -point DFT, we obtain the following N/4-point DFTs: 

N/4−1

X(4k + 1) =

{[x(n) − x(n + N/2)]

(1.56)

n=0 kn − j [x(n + N/4) − x(n + 3N/4)]}WNn WN/4



N/4−1

X(4k + 3) =

{[x(n) − x(n + N/2)]

(1.57)

n=0 kn + j [x(n + N/4) − x(n + 3N/4)]}WN3n WN/4

Thus the N -point DFT is decomposed into one N/2-point DFT without additional phase factors and two N/4-point DFTs with phase factors. The N -point DFT is obtained by successive use of these decompositions up to the last stage. Thus we obtain a decimation-in-frequency SRFFT algorithm. Figure 1.15 shows the flow graph for an in-place 32-point decimation-in-frequency SRFFT algorithm. At stage A of the computation for N = 32, the top 16 points constitute the sequence g0 (n) = x(n) + x(n + N/2),

0 ≤ n ≤ 15

(1.58)

This is the sequence required for the computation of X(2k). The next 8 points constitute the sequence g1 (n) = x(n) − x(n + N/2),

0≤n≤7

(1.59)

The bottom eight points constitute the sequence jg2 (n), where g2 (n) = x(n + N/4) − x(n + 3N/4),

0≤n≤7

(1.60)

The sequences g1 (n) and g2 (n) are used in the computation of X(4k+1) and X(4k+3). Thus, at stage A we have completed the first decimation for the radix-2 component of the algorithm. At stage B, the bottom eight points constitute the computation of 3n [g1 (n) + jg2 (n)]W32 , 0 ≤ n ≤ 7, which is used to compute X(4k + 3), 0 ≤ k ≤ 7. The next eight points from the bottom constitute the computation of [g1 (n) − jg2 (n)] n W32 , 0 ≤ n ≤ 7, which is used to compute X(4k + 1), 0 ≤ k ≤ 7. Thus at stage B, we have completed the first decimation for the radix-4 algorithm, which results in two 8-point sequences. Hence the basic butterfly computation for the SRFFT algorithm has the “L-shaped” form illustrated in Fig. 1.16. Now we repeat the steps in the computation above. Beginning with the top 16 points at stage A, we repeat the decomposition for the 16-point DFT. In other words,

545

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

j

W6W4W2

j j

j j

j

W W 6 W 1 12 8

j j j

j

j

W6W5 W4 W3 W2 W1

j j j j j

4 W W1 2

j j

4 W W1 2

7 W

W3 W6 W9 W1 2 W1 5 W1 8 W2 1

j j j

A

j j j

4 W W 12

0 16 8 24 4 20 12 28 2 18 10 26 6 22 14 30 1 17 9 25 5 21 13 29 3 19 11 27 7 23 15 31

B

Figure 1.15 Length 32 split-radix FFT algorithms from paper by Duhamel (1986); reprinted with

permission from the IEEE.

we decompose the computation into an eight-point, radix-2 DFT and two four-point, radix-4 DFTs. Thus at stage B, the top eight points constitute the sequence (with N = 16) g0 (n) = g0 (n) + g0 (n + N/2),

0≤n≤7

(1.61)

and the next eight points constitute the two four-point sequences g1 (n) and jg2 (n), where 0≤n≤3 g1 (n) = g0 (n) − g0 (n + N/2), (1.62) g2 (n) = g0 (n + N/4) − g0 (n + 3N/4), 0≤n≤3

546

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

x(n) Use for X(2k)

(

x n+

N 4

)

n

WN

(

N 2

(

3N 2

x n+

x n+

) )

−1

−j

Use for X(4k + 1)

−1

j

Use for X(4k + 3) W 3n N

Figure 1.16 Butterfly for SRFFT algorithm.

The bottom 16 points of stage B are in the form of two eight-point DFTs. Hence each eight-point DFT is decomposed into a four-point, radix-2 DFT and a four-point, radix-4 DFT. In the final stage, the computations involve the combination of twopoint sequences. Table 2 presents a comparison of the number of nontrivial real multiplications and additions required to perform an N -point DFT with complex-valued data, using a radix-2, radix-4, radix-8, and a split-radix FFT. Note that the SRFFT algorithm requires the lowest number of multiplication and additions. For this reason, it is preferable in many practical applications. Another type of SRFFT algorithm has been developed by Price (1990). Its relation to Duhamel’s algorithm described previously can be seen by noting that the radix-4 DFT terms X(4k + 1) and X(4k + 3) involve the N/4-point DFTs of the sequences [g1 (n) − jg2 (n)]WNn and [g1 (n) + jg2 (n)]WN3n , respectively. In effect, the 8 sequences g1 (n) and g2 (n) are multiplied by the factor (vector) (1, −j ) = (1, W32 ) Number of Nontrivial Real Multiplications and Additions to Compute an N -point Complex DFT

TABLE 2

Real Multiplications N 16 32 64 128 256 512 1,024

Radix 2 24 88 264 712 1,800 4,360 10,248

Radix Radix Split 4 8 Radix 20 208

204

1,392 3,204 7,856

20 68 196 516 1,284 3,076 7,172

Real Additions Radix 2 152 408 1,032 2,504 5,896 13,566 30,728

Radix 4

Radix Split 8 Radix

148 976

972

5,488 12,420 28,336

148 388 964 2,308 5,380 12,292 27,652

Source: Extracted from Duhamel (1986).

547

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

and by WNn for the computation of X(4k + 1), while the computation of X(4k + 3) −8 involves the factor (1, j ) = (1, W32 ) and WN3n . Instead, one can rearrange the −8 computation so that the factor for X(4k + 3) is (−j, −1) = −(W32 , 1). As a result of this phase rotation, the phase factors in the computation of X(4k + 3) become exactly the same as those for X(4k + 1), except that they occur in mirror image order. For example, at stage B of Fig. 1.15, the phase factors W 21, W 18 , . . . , W 3 are replaced by W 1 , W 2 , . . . , W 7 , respectively. This mirror-image symmetry occurs at every subsequent stage of the algorithm. As a consequence, the number of phase factors that must be computed and stored is reduced by a factor of 2 in comparison to Duhamel’s algorithm. The resulting algorithm is called the “mirror” FFT (MFFT) algorithm. An additional factor-of-2 savings in storage of phase factors can be obtained by introducing a 90◦ phase offset at the midpoint of each factor array, which can be removed if necessary at the output of the SRFFT computation. The incorporation of this improvement into the SRFFT (or the MFFT) results in another algorithm, also due to Price (1990), called the “phase” FFT (PFFT) algorithm.

1.6 Implementation of FFT Algorithms Now that we have described the basic radix-2 and radix-4 FFT algorithms, let us consider some of the implementation issues. Our remarks apply directly to radix-2 algorithms, although similar comments may be made about radix-4 and higher-radix algorithms. Basically, the radix-2 FFT algorithm consists of taking two data points at a time from memory, performing the butterfly computations and returning the resulting numbers to memory. This procedure is repeated many times ((N log2 N )/2 times) in the computation of an N -point DFT. The butterfly computations require the phase factors {WNk } at various stages in either natural or bit-reversed order. In an efficient implementation of the algorithm, the phase factors are computed once and stored in a table, either in normal order or in bit-reversed order, depending on the specific implementation of the algorithm. Memory requirement is another factor that must be considered. If the computations are performed in place, the number of memory locations required is 2N since the numbers are complex. However, we can instead double the memory to 4N, thus simplifying the indexing and control operations in the FFT algorithms. In this case we simply alternate in the use of the two sets of memory locations from one stage of the FFT algorithm to the other. Doubling of the memory also allows us to have both the input sequence and the output sequence in normal order. There are a number of other implementation issues regarding indexing, bit reversal, and the degree of parallelism in the computations. To a large extent, these issues are a function of the specific algorithm and the type of implementation, namely, a hardware or software implementation. In implementations based on a fixed-point arithmetic, or floating-point arithmetic on small machines, there is also the issue of round-off errors in the computation. This topic is considered in Section 4.

548

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Although the FFT algorithms described previously were presented in the context of computing the DFT efficiently, they can also be used to compute the IDFT, which is

x(n) =

N−1 1  X(k)WN−nk N

(1.63)

k=0

The only difference between the two transforms is the normalization factor 1/N and the sign of the phase factor WN . Consequently, an FFT algorithm for computing the DFT can be converted to an FFT algorithm for computing the IDFT by changing the sign on all the phase factors and dividing the final output of the algorithm by N . In fact, if we take the decimation-in-time algorithm that we described in Section 1.3, reverse the direction of the flow graph, change the sign on the phase factors, interchange the output and input, and finally, divide the output by N , we obtain a decimation-in-frequency FFT algorithm for computing the IDFT. On the other hand, if we begin with the decimation-in-frequency FFT algorithm described in Section 1.3 and repeat the changes described above, we obtain a decimation-intime FFT algorithm for computing the IDFT. Thus it is a simple matter to devise FFT algorithms for computing the IDFT. Finally, we note that the emphasis in our discussion of FFT algorithms was on radix-2, radix-4, and split-radix algorithms. These are by far the most widely used in practice. When the number of data points is not a power of 2 or 4, it is a simple matter to pad the sequence x(n) with zeros such that N = 2ν or N = 4ν . The measure of complexity for FFT algorithms that we have emphasized is the required number of arithmetic operations (multiplications and additions). Although this is a very important benchmark for computational complexity, there are other issues to be considered in practical implementation of FFT algorithms. These include the architecture of the processor, the available instruction set, the data structures for storing phase factors, and other considerations. For general-purpose computers, where the cost of the numerical operations dominates, radix-2, radix-4, and split-radix FFT algorithms are good candidates. However, in the case of special-purpose digital signal processors, featuring single-cycle multiply-and-accumulate operation, bit-reversed addressing, and a high degree of instruction parallelism, the structural regularity of the algorithm is equally as important as arithmetic complexity. Hence for DSP processors, radix-2 or radix-4 decimationin-frequency FFT algorithms are preferable in terms of speed and accuracy. The irregular structure of the SRFFT may render it less suitable for implementation on digital signal processors. Structural regularity is also important in the implementation of FFT algorithms on vector processors, multiprocessors, and in VLSI. Interprocessor communication is an important consideration in such implementations on parallel processors. In conclusion, we have presented several important considerations in the implementation of FFT algorithms. Advances in digital signal processing technology, in hardware and software, will continue to influence the choice among FFT algorithms for various practical applications.

549

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

2

Applications of FFT Algorithms The FFT algorithms described in the preceding section find application in a variety of areas, including linear filtering, correlation, and spectrum analysis. Basically, the FFT algorithm is used as an efficient means to compute the DFT and the IDFT. In this section we consider the use of the FFT algorithm in linear filtering and in the computation of the crosscorrelation of two sequences. In addition we illustrate how to enhance the efficiency of the FFT algorithm by forming complex-valued sequences from real-valued sequences prior to the computation of the DFT.

2.1

Efficient Computation of the DFT of Two Real Sequences

The FFT algorithm is designed to perform complex multiplications and additions, even though the input data may be real valued. The basic reason for this situation is that the phase factors are complex and hence, after the first stage of the algorithm, all variables are basically complex valued. In view of the fact that the algorithm can handle complex-valued input sequences, we can exploit this capability in the computation of the DFT of two real-valued sequences. Suppose that x1 (n) and x2 (n) are two real-valued sequences of length N , and let x(n) be a complex-valued sequence defined as x(n) = x1 (n) + j x2 (n),

0≤n≤N −1

(2.1)

The DFT operation is linear and hence the DFT of x(n) can be expressed as X(k) = X1 (k) + j X2 (k)

(2.2)

The sequences x1 (n) and x2 (n) can be expressed in terms of x(n) as follows: x1 (n) =

x(n) + x ∗ (n) 2

(2.3)

x2 (n) =

x(n) − x ∗ (n) 2j

(2.4)

X1 (k) =

1 {DFT[x(n)] + DFT[x ∗ (n)]} 2

(2.5)

X2 (k) =

1 {DFT[x(n)] − DFT[x ∗ (n)]} 2j

(2.6)

Hence the DFTs of x1 (n) and x2 (n) are

Recall that the DFT of x ∗ (n) is X ∗ (N − k). Therefore,

550

X1 (k) =

1 [X(k) + X ∗ (N − k)] 2

(2.7)

X2 (k) =

1 [X(k) − X ∗ (N − k)] j2

( 2.8)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Thus, by performing a single DFT on the complex-valued sequence x(n), we have obtained the DFT of the two real sequences with only a small amount of additional computation that is involved in computing X1 (k) and X2 (k) from X(k) by use of (2.7) and (2.8).

2.2

Efficient Computation of the DFT of a 2 N -Point Real Sequence

Suppose that g(n) is a real-valued sequence of 2N points. We now demonstrate how to obtain the 2N -point DFT of g(n) from computation of one N -point DFT involving complex-valued data. First, we define x1 (n) = g(2n)

(2.9)

x2 (n) = g(2n + 1)

Thus we have subdivided the 2N -point real sequence into two N -point real sequences. Now we can apply the method described in the preceding section. Let x(n) be the N -point complex-valued sequence x(n) = x1 (n) + j x2 (n)

(2.10)

From the results of the preceding section, we have X1 (k) =

1 [X(k) + X ∗ (N − k)] 2

1 X2 (k) = [X(k) − X ∗ (N − k)] 2j

(2.11)

Finally, we must express the 2N -point DFT in terms of the two N -point DFTs, X1 (k) and X2 (k). To accomplish this, we proceed as in the decimation-in-time FFT algorithm, namely, G(k) =

N−1 

2nk g(2n)W2N +

n=0

=

N−1 

N−1 

(2n+1)k g(2n + 1)W2N

n=0

k x1 (n)WNnk + W2N

n=0

N −1 

x2 (n)WNnk

n=0

Consequently, G(k) = X1 (k) + W2k N X2 (k),

k = 0, 1, . . . , N − 1

G(k + N ) = X1 (k) − W2k N X2 (k),

k = 0, 1, . . . , N − 1

(2.12)

Thus we have computed the DFT of a 2N -point real sequence from one N -point DFT and some additional computation as indicated by (2.11) and (2.12).

551

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

2.3

Use of the FFT Algorithm in Linear Filtering and Correlation

An important application of the FFT algorithm is in FIR linear filtering of long data sequences. The two methods are the overlap-add and the overlapsave methods for filtering a long data sequence with an FIR filter, based on the use of the DFT. In this section we consider the use of these two methods in conjunction with the FFT algorithm for computing the DFT and the IDFT. Let h(n), 0 ≤ n ≤ M − 1, be the unit sample response of the FIR filter and let x(n) denote the input data sequence. The block size of the FFT algorithm is N , where N = L + M − 1 and L is the number of new data samples being processed by the filter. We assume that for any given value of M , the number L of data samples is selected so that N is a power of 2. For purposes of this discussion, we consider only radix-2 FFT algorithms. The N -point DFT of h(n), which is padded by L − 1 zeros, is denoted as H (k). This computation is performed once via the FFT and the resulting N complex numbers are stored. To be specific we assume that the decimation-in-frequency FFT algorithm is used to compute H (k). This yields H (k) in bit-reversed order, which is the way it is stored in memory. In the overlap-save method, the first M − 1 data points of each data block are the last M − 1 data points of the previous data block. Each data block contains L new data points, such that N = L + M − 1. The N -point DFT of each data block is performed by the FFT algorithm. If the decimation-in-frequency algorithm is employed, the input data block requires no shuffling and the values of the DFT occur in bit-reversed order. Since this is exactly the order of H (k), we can multiply the DFT of the data, say Xm (k), with H (k), and thus the result Ym (k) = H (k)Xm (k) is also in bit-reversed order. The inverse DFT (IDFT) can be computed by use of an FFT algorithm that takes the input in bit-reversed order and produces an output in normal order. Thus there is no need to shuffle any block of data in computing either the DFT or the IDFT. If the overlap-add method is used to perform the linear filtering, the computational method using the FFT algorithm is basically the same. The only difference is that the N -point data blocks consist of L new data points and M − 1 additional zeros. After the IDFT is computed for each data block, the N -point filtered blocks are overlapped, and the M − 1 overlapping data points between successive output records are added together. Let us assess the computational complexity of the FFT method for linear filtering. For this purpose, the one-time computation of H (k) is insignificant and can be ignored. Each FFT requires (N/2) log2 N complex multiplications and N log2 N additions. Since the FFT is performed twice, once for the DFT and once for the IDFT, the computational burden is N log2 N complex multiplications and 2N log2 N additions. There are also N complex multiplications and N − 1 additions required to compute Ym (k). Therefore, we have (N log2 2N )/L complex multiplications per output data point and approximately (2N log2 2N )/L additions per output data point.

552

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

The overlap-add method requires an incremental increase of (M − 1)/L in the number of additions. By way of comparison, a direct-form realization of the FIR filter involves M real multiplications per output point if the filter is not linear phase, and M/2 if it is linear phase (symmetric). Also, the number of additions is M − 1 per output point. It is interesting to compare the efficiency of the FFT algorithm with the direct form realization of the FIR filter. Let us focus on the number of multiplications, which are more time consuming than additions. Suppose that M = 128 = 27 and N = 2ν . Then the number of complex multiplications per output point for an FFT size of N = 2ν is c(ν) = ≈

N log2 2N 2ν (ν + 1) = L N −M +1 2ν (ν + 1) 2ν − 27

The values of c(ν) for different values of ν are given in Table 3. We observe that there is an optimum value of ν which minimizes c(ν). For the FIR filter of size M = 128, the optimum occurs at ν = 10. We should emphasize that c(ν) represents the number of complex multiplications for the FFT-based method. The number of real multiplications is four times this number. However, even if the FIR filter has linear phase, the number of computa tions per output point is still less with the FFT-based method. Further- more, the efficiency of the FFT method can be improved by computing the DFT of two succ essive data blocks simultaneously, according to the method just described. Conse quently, the FFT-based method is indeed superior from a computational point of view when the filter length is relatively large. The computation of the cross correlation between two sequences by means of the FFT algorithm is similar to the linear FIR filtering problem just described. In practical applications involving crosscorrelation, at least one of the sequences has finite duration and is akin to the impulse response of the FIR filter. The second sequence may be a long sequence which contains the desired sequence corrupted by additive noise. Hence the second sequence is akin to the input to the FIR filter. TABLE 3 Computational Complexity

Size of FFT ν = log2 N

c(ν) Number of Complex Multiplications per Output Point

9

13.3

10

12.6

11

12.8

12

13.4

14

15.1

553

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

By time reversing the first sequence and computing its DFT, we have reduced the cross correlation to an equivalent convolution problem (i.e., a linear FIR filtering problem). Therefore, the methodology we developed for linear FIR filtering by use of the FFT applies directly.

3

A Linear Filtering Approach to Computation of the DFT The FFT algorithm takes N points of input data and produces an output sequence of N points corresponding to the DFT of the input data. As we have shown, the radix-2 FFT algorithm performs the computation of the DFT in (N/2) log2 N multiplications and N log2 N additions for an N -point sequence. There are some applications where only a selected number of values of the DFT are desired, but the entire DFT is not required. In such a case, the FFT algorithm may no longer be more efficient than a direct computation of the desired values of the DFT. In fact, when the desired number of values of the DFT is less than log2 N , a direct computation of the desired values is more efficient. The direct computation of the DFT can be formulated as a linear filtering operation on the input data sequence. As we will demonstrate, the linear filter takes the form of a parallel bank of resonators where each resonator selects one of the frequencies ωk = 2πk/N , k = 0, 1, . . . , N − 1, corresponding to the N frequencies in the DFT. There are other applications in which we require the evaluation of the z-transform of a finite-duration sequence at points other than the unit circle. If the set of desired points in the z-plane possesses some regularity, it is possible to also express the computation of the z-transform as a linear filtering operation. In this connection, we introduce another algorithm, called the chirp-z transform algorithm, which is suitable for evaluating the z-transform of a set of data on a variety of contours in the z-plane. This algorithm is also formulated as a linear filtering of a set of input data. As a consequence, the FFT algorithm can be used to compute the chirp-z transform and thus to evaluate the z-transform at various contours in the z-plane, including the unit circle.

3.1

The Goertzel Algorithm

The Goertzel algorithm exploits the periodicity of the phase factors { WNk } and allows us to express the computation of the DFT as a linear filtering operation. Since WN−kN = 1, we can multiply the DFT by this factor. Thus X(k) = WN−kN

N−1 

x(m)WNkm =

m=0

N−1 

x(m)WN−k(N−m)

(3.1)

m=0

We note that (3.1) is in the form of a convolution. Indeed, if we define the sequence yk (n) as N−1  (3.2) x(m)WN−k(n−m) yk (n) = m=0

554

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

then it is clear that yk (n) is the convolution of the finite-duration input sequence x(n) of length N with a filter that has an impulse response hk (n) = WN−kn u(n)

(3.3)

The output of this filter at n = N yields the value of the DFT at the frequency ωk = 2πk/N . That is, X(k) = yk (n)|n=N

(3.4)

as can be verified by comparing (3.1) with (3.2). The filter with impulse response hk (n) has the system function Hk (z) =

1

(3.5)

1 − WN−k z−1

This filter has a pole on the unit circle at the frequency ωk = 2π k/N . Thus, the entire DFT can be computed by passing the block of input data into a parallel bank of N single-pole filters (resonators), where each filter has a pole at the corresponding frequency of the DFT. Instead of performing the computation of the DFT as in (3.2), via convolution, we can use the difference equation corresponding to the filter given by (3.5) to compute yk (n) recursively. Thus we have yk (n) = WN−k yk (n − 1) + x(n),

yk (−1) = 0

(3.6)

The desired output is X(k) = yk (N ), for k = 0, 1, . . . , N − 1. To perform this computation, we can compute once and store the phase factors WN−k . The complex multiplications and additions inherent in (3.6) can be avoided by combining the pairs of resonators possessing complex-conjugate poles. This leads to two-pole filters with system functions of the form Hk (z) =

1 − WNk z−1 1 − 2 cos(2π k/N )z−1 + z−2

(3.7)

The direct form II realization of the system illustrated in Fig. 3.1 is described by the difference equations vk (n) = 2 cos

2π k vk (n − 1) − vk (n − 2) + x(n) N

yk (n) = vk (n) − WNk vk (n − 1)

(3.8) (3.9)

with initial conditions vk (−1) = vk (−2) = 0.

555

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

x(n)

vk(n)

+

+

yk(n)

z −1 + 2 cos 2πk N

Figure 3.1

Direct form II realization of two-pole resonator for computing the DFT.

W nk z −1

−1

The recursive relation in (3.8) is iterated for n = 0, 1, . . . , N , but the equation in (3.9) is computed only once at time n = N . Each iteration requires one real multiplication and two additions. Consequently, for a real input sequence x(n), this algorithm requires N + 1 real multiplications to yield not only X(k) but also, due to symmetry, the value of X(N − k). The Goertzel algorithm is particularly attractive when the DFT is to be computed at a relatively small number M of values, where M ≤ log2 N . Otherwise, the FFT algorithm is a more efficient method.

3.2

The Chirp- z Transform Algorithm

The DFT of an N -point data sequence x(n) has been viewed as the z-transform of x(n) evaluated at N equally spaced points on the unit circle. It has also been viewed as N equally spaced samples of the Fourier transform of the data sequence x(n). In this section we consider the evaluation of X(z) on other contours in the z-plane, including the unit circle. Suppose that we wish to compute the values of the z-transform of x(n) at a set of points {zk }. Then, X(zk ) =

N−1 

x(n)zk−n ,

k = 0, 1, . . . , L − 1

(3.10)

n=0

For example, if the contour is a circle of radius r and the zk are N equally spaced points, then

X(zk ) =

zk = rej 2πkn/N ,

k = 0, 1, 2, . . . , N − 1

[x(n)r −n ]e−j 2πkn/N ,

k = 0, 1, 2, . . . , N − 1

N−1 

(3.11)

n=0

In this case the FFT algorithm can be applied on the modified sequence x(n)r −n . More generally, suppose that the points zk in the z-plane fall on an arc which begins at some point z0 = r0 ej θ0

556

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

and spirals either in toward the origin or out away from the origin such that the points {zk } are defined as zk = r0 ej θ0 (R0 ej φ0 )k ,

k = 0, 1, . . . , L − 1

(3.12)

Note that if R0 < 1, the points fall on a contour that spirals toward the origin, and if R0 > 1, the contour spirals away from the origin. If R0 = 1, the contour is a circular arc of radius r0 . If r0 = 1 and R0 = 1, the contour is an arc of the unit circle. The latter contour would allow us to compute the frequency content of the sequence x(n) at a dense set of L frequencies in the range covered by the arc without having to compute a large DFT, that is, a DFT of the sequence x(n) padded with many zeros to obtain the desired resolution in frequency. Finally, if r0 = R0 = 1, θ0 = 0, φ0 = 2π/N , and L = N , the contour is the entire unit circle and the frequencies are those of the DFT. The various contours are illustrated in Fig. 3.2. Im(z)

Im(z) Unit circle

Unit circle r0 Re(z)

Re(z)

R0 = 1 r0 < 1 φ0 = θ0 = 0

R0 = r0 = 1 φ0 = θ0 = 0

Im(z)

Im(z) Unit circle

r0

R0 < 1

θ0

Unit circle

Re(z)

Re(z)

R0 > 1

Figure 3.2 Some examples of contours on which we may evaluate the z-

transform.

557

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

When points {zk } in (3.12) are substituted into the expression for the z-transform, we obtain N−1  x(n)zk−n X(zk ) = n=0

=

N−1 

(3.13) x(n)(r0 ej θ0 )−n V −nk

n=0

where, by definition, V = R0 ej φ0

(3.14)

We can express (3.13) in the form of a convolution, by noting that nk = 21 [n2 + k 2 − (k − n)2 ]

(3.15)

Substitution of (3.15) into (3.13) yields X(zk ) = V

−k 2 /2

N−1 

[x(n)(r0 ej θ0 )−n V −n

2 /2

]V (k−n)

2 /2

(3.16)

n=0

Let us define a new sequence g(n) as g(n) = x(n)(r0 ej θ0 )−n V −n

2 /2

(3.17)

Then (3.16) can be expressed as X(zk ) = V −k

2 /2

N−1 

g(n)V (k−n)

2 /2

(3.18)

n=0

The summation in (3.18) can be interpreted as the convolution of the sequence g(n) with the impulse response h(n) of a filter, where h(n) = V n

2 /2

(3.19)

Consequently, (3.18) may be expressed as X(zk ) = V −k =

2 /2

y(k) , h(k)

y(k) (3.20) k = 0, 1, . . . , L − 1

where y(k) is the output of the filter y(k) =

N−1  n=0

558

g(n)h(k − n),

k = 0, 1, . . . , L − 1

(3.21)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

We observe that both h(n) and g(n) are complex-valued sequences. The sequence h(n) with R0 = 1 has the form of a complex exponential with argument ωn = n2 φ0 /2 = (nφ0 /2)n. The quantity nφ0 /2 represents the frequency of the complex exponential signal, which increases linearly with time. Such signals are used in radar systems and are called chirp signals. Hence the z-transform evaluated as in (3.18) is called the chirp- z transform. The linear convolution in (3.21) is most efficiently done by use of the FFT algorithm. The sequence g(n) is of length N . However, h(n) has infinite duration. Fortunately, only a portion h(n) is required to compute the L values of X(z). Since we will compute the convolution in (3.21) via the FFT, let us consider the circular convolution of the N -point sequence g(n) with an M -point section of h(n), where M > N . In such a case, we know that the first N − 1 points contain aliasing and that the remaining M − N + 1 points are identical to the result that would be obtained from a linear convolution of h(n) with g(n). In view of this, we should select a DFT of size M =L+N −1 which would yield L valid points and N − 1 points corrupted by aliasing. The section of h(n) that is needed for this computation corresponds to the values of h(n) for −(N − 1) ≤ n ≤ (L − 1), which is of length M = L + N − 1, as observed from (3.21). Let us define the sequence h (n) 1 of length M as h1 (n) = h(n − N + 1),

n = 0, 1, . . . , M − 1

(3.22)

and compute its M -point DFT via the FFT algorithm to obtain H1 (k). From x(n) we compute g(n) as specified by (3.17), pad g(n) with L − 1 zeros, and compute its M -point DFT to yield G(k). The IDFT of the product Y1 (k) = G(k)H1 (k) yields the M -point sequence y1 (n), n = 0, 1, . . . , M − 1. The first N − 1 points of y1 (n) are corrupted by aliasing and are discarded. The desired values are y1 (n) for N − 1 ≤ n ≤ M − 1, which correspond to the range 0 ≤ n ≤ L − 1 in (3.21), that is, y(n) = y1 (n + N − 1),

n = 0, 1, . . . , L − 1

(3.23)

Alternatively, we can define a sequence h2 (n) as

h2 (n) =

h(n), h(n − N − L + 1),

0≤n≤L−1 L≤n≤M −1

(3.24)

The M -point DFT of h2 (n) yields H2 (k), which when multiplied by G(k) yields Y2 (k) = G(k)H2 (k). The IDFT of Y2 (k) yields the sequence y2 (n) for 0 ≤ n ≤ M − 1. Now the desired values of y2 (n) are in the range 0 ≤ n ≤ L − 1, that is, y(n) = y2 (n),

n = 0, 1, . . . , L − 1

(3.25)

Finally, the complex values X(zk ) are computed by dividing y(k) by h(k), k = 0, 1, . . . , L − 1, as specified by (3.20).

559

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

In general, the computational complexity of the chirp-z transform algorithm described above is of the order of M log2 M complex multiplications, where M = N + L − 1. This number should be compared with the product, N · L, the number of computations required by direct evaluation of the z-transform. Clearly, if L is small, direct computation is more efficient. However, if L is large, then the chirp-z transform algorithm is more efficient. The chirp-z transform method has been implemented in hardware to compute the DFT of signals. For the computation of the DFT, we select r0 = R0 = 1, θ0 = 0, φ0 = 2π/N, and L = N. In this case V −n

2 /2

= e−j πn

2 /N

(3.26)

π n2 π n2 − j sin = cos N N The chirp filter with impulse response h(n) = V n

2 /2

= cos

π n2 π n2 + j sin N N

(3.27)

= hr (n) + j hi (n) has been implemented as a pair of FIR filters with coefficients hr (n) and hi (n), respectively. Both surface acoustic wave (SAW) devices and charge coupled devices

ROM cos πn N

2

FIR filter 2 hr (n) = cos πn N

+ −

( )2

+

x FIR filter 2 hi (n) = sin πn N

x(n)

+



FIR filter 2 hi (n) = sin πn N x − sin

πn2 N ROM

FIR filter 2 hr (n) = cos πn N

+

+ +

( )2

Chirp Filters

Figure 3.3 Block diagram illustrating the implementation of the chirp-z transform

for computing the DFT (magnitude only).

560

y(n)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

(CCD) have been used in practice for the FIR filters. The cosine and sine sequences given in (3.26) needed for the premultiplications and postmultiplications are usually stored in a read-only memory (ROM). Furthermore, we note that if only the magnitude of the DFT is desired, the postmultiplications are unnecessary. In this case, |X(zk )| = |y(k)|, k = 0, 1, . . . , N − 1 (3.28) as illustrated in Fig. 3.3. Thus the linear FIR filtering approach using the chirp-z transform has been implemented for the computation of the DFT.

4

Quantization Effects in the Computation of the DFT 1 As we have observed in our previous discussions, the DFT plays an important role in many digital signal processing applications, including FIR filtering, the computation of the correlation between signals, and spectral analysis. For this reason it is important for us to know the effect of quantization errors in its computation. In particular, we shall consider the effect of round-off errors due to the multiplications performed in the DFT with fixed-point arithmetic. The model that we shall adopt for characterizing round-off errors in multiplication is the additive white noise model thatweuse in the statistical analysis of round-off errors in IIR and FIR filters. Although the statistical analysis is performed for rounding, the analysis can be easily modified to apply to truncation in two’s-complement arithmetic. Of particular interest is the analysis of round-off errors in the computation of the DFT via the FFT algorithm. However, we shall first establish a benchmark by determining the round-off errors in the direct computation of the DFT.

4.1

Quantization Errors in the Direct Computation of the DFT

Given a finite-duration sequence {x(n)}, 0 ≤ n ≤ N − 1, the DFT of {x(n)} is defined as N−1  x(n)WNkn , k = 0, 1, . . . , N − 1 (4.1) X(k) = n=0 −j 2π/N

. We assume that in general, {x(n)} is a complex-valued sewhere WN = e quence. We also assume that the real and imaginary components of {x(n)} and {WNkn } are represented by b bits. Consequently, the computation of the product x(n)WNkn requires four real multiplications. Each real multiplication is rounded from 2b bits to b bits, and hence there are four quantization errors for each complex-valued multiplication. In the direct computation of the DFT, there are N complex-valued multiplications for each point in the DFT. Therefore, the total number of real multiplications in the computation of a single point in the DFT is 4N . Consequently, there are 4N quantization errors. 1

It is recommended that the reader review quantization of filter coefficients prior to reading this section.

561

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Let us evaluate the variance of the quantization errors in a fixed-point computation of the DFT. First, we make the following assumptions about the statistical properties of the quantization errors. 1. The quantization errors due to rounding are uniformly distributed random variables in the range (−/2, /2) where  = 2−b . 2. The 4N quantization errors are mutually uncorrelated. 3. The 4N quantization errors are uncorrelated with the sequence {x(n)}. Since each of the quantization errors has a variance 2 2−2b = 12 12

σe2 =

(4.2)

the variance of the quantization errors from the 4N multiplications is σq2 = 4N σe2 =

N −2b ·2 3

(4.3)

Hence the variance of the quantization error is proportional to the size of DFT. Note that when N is a power of 2 (i.e., N = 2ν ), the variance can be expressed as σq2 =

2−2(b−ν/2) 3

(4.4)

This expression implies that every fourfold increase in the size N of the DFT requires an additional bit in computational precision to offset the additional quantization errors. To prevent overflow, the input sequence to the DFT requires scaling. Clearly, an upper bound on |X(k)| is N−1  |x(n)| |X(k)| ≤ (4.5) n=0

If the dynamic range in addition is (−1, 1), then |X(k)| < 1 requires that N−1 

|x(n)| < 1

(4.6)

n=0

If |x(n)| is initially scaled such that |x(n)| < 1 for all n, then each point in the sequence can be divided by N to ensure that (4.6) is satisfied. The scaling implied by (4.6) is extremely severe. For example, suppose that the signal sequence {x(n)} is white and, after scaling, each value |x(n)| of the sequence is uniformly distributed in the range (−1/N, 1/N ). Then the variance of the signal sequence is (2/N )2 1 = σx2 = (4.7) 12 3N 2

562

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

and the variance of the output DFT coefficients |X(k)| is σX2 = N σx2 =

(4.8)

1 3N

Thus the signal-to-noise power ratio is σX2 22b = σq2 N2

(4.9)

We observe that the scaling is responsible for reducing the SNR by N and the combination of scaling and quantization errors results in a total reduction that is proportional to N 2 . Hence scaling the input sequence {x(n)} to satisfy (4.6) imposes a severe penalty on the signal-to-noise ratio in the DFT. EXAMPLE 4.1 Use (4.9) to determine the number of bits required to compute the DFT of a 1024-point sequence with an SNR of 30 dB. Solution.

The size of the sequence is N = 210 . Hence the SNR is 10 log10

σX2 = 10 log10 22b−20 σq2

For an SNR of 30 dB, we have 3(2b − 20) = 30 b = 15 bits Note that the 15 bits is the precision for both multiplication and addition.

Instead of scaling the input sequence {x(n)}, suppose we simply require that |x(n)| < 1. Then we must provide a sufficiently large dynamic range for addition such that |X(k)| < N . In such a case, the variance of the sequence {|x(n)|} is σx2 = 13 , and hence the variance of |X(k)| is σX2 = N σx2 =

N 3

(4.10)

Consequently, the SNR is σX2 = 22b σq2

(4.11)

If we repeat the computation in Example 4.1, we find that the number of bits required to achieve an SNR of 30 dB is b = 5 bits. However, we need an additional 10 bits for the accumulator (the adder) to accommodate the increase in the dynamic range for addition. Although we did not achieve any reduction in the dynamic range for addition, we have managed to reduce the precision in multiplication from 15 bits to 5 bits, which is highly significant.

563

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

4.2 Quantization Errors in FFT Algorithms As we have shown, the FFT algorithms require significantly fewer multiplications than the direct computation of the DFT. In view of this we might conclude that the computation of the DFT via an FFT algorithm will result in smaller quantization errors. Unfortunately, that is not the case, as we will demonstrate. Let us consider the use of fixed-point arithmetic in the computation of a radix-2 FFT algorithm. To be specific, we select the radix-2, decimation-in-time algorithm illustrated in Fig. 4.1 for the case N = 8. The results on quantization errors that we obtain for this radix-2 FFT algorithm are typical of the results obtained with other radix-2 and higher radix algorithms. We observe that each butterfly computation involves one complex-valued multiplication or, equivalently, four real multiplications. We ignore the fact that some butterflies contain a trivial multiplication by ±1. If we consider the butterflies that affect the computation of any one value of the DFT, we find that, in general, there are N/2 in the first stage of the FFT, N/4 in the second stage, N/8 in the third state, and so on, until the last stage, where there is only one. Consequently, the number of Stage 1

Stage 2

Stage 3 X(0)

x(0)

x(4)

W 08

W 08

x(2)

x(6)

X(1)

−1

W 08

W 28

X(3)

−1

−1

W 08

x(1)

−1 0

x(5)

W 18

W8

−1

−1 W 08

x(3)

2

W8

−1

Figure 4.1 Decimation-in-time FFT algorithm.

564

−1 W 38

W8 −1

X(4)

X(5)

2

W8 −1

0

x(7)

X(2)

−1

−1

X(6)

X(7)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

butterflies per output point is     ν−1  2ν−1 + 2ν−2 + · · · + 2 + 1 = 2ν−1 1 + 21 + · · · + 21 = 2 [1 − ν

 ν 1 2

(4.12) ]=N −1

For example, the butterflies that affect the computation of X(3) in the eight-point FFT algorithm of Fig. 4.1 are illustrated in Fig. 4.2. The quantization errors introduced in each butterfly propagate to the output. Note that the quantization errors introduced in the first stage propagate through (ν − 1) stages, those introduced in the second stage propagate through (ν − 2) stages, and so on. As these quantization errors propagate through a number of subsequent stages, they are phase shifted (phase rotated) by the phase factors WNkn . These phase rotations do not change the statistical properties of the quantization errors and, in particular, the variance of each quantization error remains invariant. If we assume that the quantization errors in each butterfly are uncorrelated with the errors in other butterflies, then there are 4(N − 1) errors that affect the output of each point of the FFT. Consequently, the variance of the total quantization error at the output is 2 N 2 σq2 = 4(N − 1) ≈ (4.13) 12 3 x(0) 0

x(4)

W8 −1

x(2)

x(6)

W 08

W 28 −1

X(3)

−1

x(1) 0

x(5)

W8 −1

x(3)

x(7)

2

W 08

W 38

W8 −1

−1

Figure 4.2 Butterflies that affect the computation of X(3).

565

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

where  = 2−b . Hence

N −2b ·2 (4.14) 3 This is exactly the same result that we obtained for the direct computation of the DFT. The result in (4.14) should not be surprising. In fact, the FFT algorithm does not reduce the number of multiplications required to compute a single point of the DFT. It does, however, exploit the periodicities in WNkn and thus reduces the number of multiplications in the computation of the entire block of N points in the DFT. As in the case of the direct computation of the DFT, we must scale the input sequence to prevent overflow. Recall that if |x(n)| < 1/N , 0 ≤ n ≤ N − 1, then |X(k)| < 1 for 0 ≤ k ≤ N − 1. Thus overflow is avoided. With this scaling, the relations in (4.7), (4.8), and (4.9), obtained previously for the direct computation of the DFT, apply to the FFT algorithm as well. Consequently, the same SNR is obtained for the FFT. Since the FFT algorithm consists of a sequence of stages, where each stage contains butterflies that involve pairs of points, it is possible to devise a different scaling strategy that is not as severe as dividing each input point by N . This alternative scaling strategy is motivated by the observation that the intermediate values |Xn (k)| in the n = 1, 2, . . . , ν stages of the FFT algorithm satisfy the conditions (see Problem 35) σq2 =

max[|Xn+1 (k)|, |Xn+1 (l)|] ≥ max[|Xn (k)|, |Xn (l)|] max[|Xn+1 (k)|, |Xn+1 (l)|] ≤ 2max[|Xn (k)|, |Xn (l)|]

(4.15)

In view of these relations, we can distribute the total scaling of 1/N into each of the stages of the FFT algorithm. In particular, if |x(n)| < 1, we apply a scale factor of 21 in the first stage so that |x(n)| < 21 . Then the output of each subsequent stage in the FFT algorithm is scaled by 21 , so that after ν stages we have achieved an overall scale factor of ( 21 )ν = 1/N . Thus overflow in the computation of the DFT is avoided. This scaling procedure does not affect the signal level at the output of the FFT algorithm, but it significantly reduces the variance of the quantization errors at the output. Specifically, each factor of 21 reduces the variance of a quantization error term by a factor of 41 . Thus the 4(N/2) quantization errors introduced in the first stage are reduced in variance by ( 41 )ν−1 , the 4(N/4) quantization errors introduced in the second stage are reduced in variance by ( 41 )ν−2 , and so on. Consequently, the total variance of the quantization errors at the output of the FFT algorithm is       ν−2   ν−3 2 N 1 ν−1 N 1 N 1 2 σq = 4 +4 +4 + ··· + 4 12 2 4 4 4 8 4    ν−2 2 1 ν−1 1 1 = + + ··· + + 1 3 2 2 2   ν  22 2 1 = ≈ · 2−2b 1− 3 2 3 where the factor ( 21 )ν is negligible.

566

(4.16)

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

We now observe that (4.16) is no longer proportional to N . On the other hand, the signal has the variance σX2 = 1/3N , as given in (4.8). Hence the SNR is σX2 1 = · 22b σq2 2N

(4.17)

= 22b−ν−1 Thus, by distributing the scaling of 1/N uniformly throughout the FFT algorithm, we have achieved an SNR that is inversely proportional to N instead of N 2 . EXAMPLE 4.2 Determine the number of bits required to compute an FFT of 1024 points with an SNR of 30 dB when the scaling is distributed as described above. Solution.

The size of the FFT is N = 210 . Hence the SNR according to (4.17) is 10 log10 22b−ν−1 = 30 3(2b − 11) = 30 b=

21 (11 bits) 2

This can be compared with the 15 bits required if all the scaling is performed in the first stage of the FFT algorithm.

5

Summary and References The focus of this chapter was on the efficient computation of the DFT. We demonstrated that by taking advantage of the symmetry and periodicity properties of the exponential factors WNkn , we can reduce the number of complex multiplications needed to compute the DFT from N 2 to N log2 N when N is a power of 2. As we indicated, any sequence can be augmented with zeros, such that N = 2ν . For decades, FFT-type algorithms were of interest to mathematicians who were concerned with computing values of Fourier series by hand. However, it was not until Cooley and Tukey (1965) published their well-known paper that the impact and significance of the efficient computation of the DFT was recognized. Since then the Cooley–Tukey FFT algorithm and its various forms, for example, the algorithms of Singleton (1967, 1969), have had a tremendous influence on the use of the DFT in convolution, correlation, and spectrum analysis. For a historical perspective on the FFT algorithm, the reader is referred to the paper by Cooley et al. (1967). The split-radix FFT (SRFFT) algorithm described in Section 1.5 is due to Duhamel and Hollmann (1984, 1986). The “mirror” FFT (MFFT) and “phase” FFT (PFFT) algorithms were described to the authors by R. Price. The exploitation of symmetry properties in the data to reduce the computation time is described in a paper by Swarztrauber (1986).

567

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

Over the years, a number of tutorial papers have been published on FFT algorithms. We cite the early papers by Brigham and Morrow (1967), Cochran et al. (1967), Bergland (1969), and Cooley et al. (1967, 1969). The recognition that the DFT can be arranged and computed as a linear convolution is also highly significant. Goertzel (1968) indicated that the DFT can be computed via linear filtering, although the computational savings of this approach is rather modest, as we have observed. More significant is the work of Bluestein (1970), who demonstrated that the computation of the DFT can be formulated as a chirp linear filtering operation. This work led to the development of the chirp-z transform algorithm by Rabiner et al. (1969). In addition to the FFT algorithms described in this chapter, there are other efficient algorithms for computing the DFT, some of which further reduce the number of multiplications, but usually require more additions. Of particular importance is an algorithm due to Rader and Brenner (1976), the class of prime factor algorithms, such as the Good algorithm (1971), and the Winograd algorithm (1976, 1978). For a description of these and related algorithms, the reader may refer to the text by Blahut (1985).

Problems 1 Show that each of the numbers ej (2π/N)k ,

0≤k ≤N −1

corresponds to an N th root of unity. Plot these numbers as phasors in the complex plane and illustrate, by means of this figure, the orthogonality property N−1 

ej (2π/N)kn e−j (2π/N)ln =

n=0

2



N, 0,

if mod (k − l, N ) = 0 otherwise

(a) Show that the phase factors can be computed recursively by ql

q

q(l−1)

WN = WN WN

(b) Perform this computation once using single-precision floating-point arithmetic and once using only four significant digits. Note the deterioration due to the accumulation of round-off errors in the latter case. (c) Show how the results in part (b) can be improved by resetting the result to the correct value −j , each time ql = N/4. Let x(n) be a real-valued N -point (N = 2ν ) sequence. Develop a method to compute an N -point DFT X (k), which contains only the odd harmonics [i.e., X  (k) = 0 if k is even] by using only a real N/2-point DFT. 4 A designer has available a number of eight-point FFT chips. Show explicitly how he should interconnect three such chips in order to compute a 24-point DFT. 3

568

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

5 The z-transform of the sequence x(n) = u(n) − u(n − 7) is sampled at five points on the unit circle as follows: x(k) = X(z)|z = ej 2πk/5 ,

k = 0, 1, 2, 3, 4

Determine the inverse DFT x  (n) of X(k). Compare it with x(n) and explain the results. 6 Consider a finite-duration sequence x(n), 0 ≤ n ≤ 7, with z-transform X(z). We wish to compute X(z) at the following set of values: zk = 0.8ej [(2πk/8)+(π/8)] ,

0≤k≤7

(a) Sketch the points {zk } in the complex plane. (b) Determine a sequence s(n) such that its DFT provides the desired samples of X(z). 7 Derive the radix-2 decimation-in-time FFT algorithm given by (1.26) and (1.27) as a special case of the more general algorithmic procedure given by (1.16) through (1.18). 8 Compute the eight-point DFT of the sequence

1, 0 ≤ n ≤ 7 x(n) = 0, otherwise by using the decimation-in-frequency FFT algorithm described in the text. 9 Derive the signal flow graph for the N = 16-point, radix-4 decimation-in-time FFT algorithm in which the input sequence is in normal order and the computations are done in place. 10 Derive the signal flow graph for the N = 16-point, radix-4 decimation-in-frequency FFT algorithm in which the input sequence is in digit-reversed order and the output DFT is in normal order. 11 Compute the eight-point DFT of the sequence

1 1 1 1 x(n) = , , , , 0, 0, 0, 0 2 2 2 2 using the in-place radix-2 decimation-in-time and radix-2 decimation-in-frequency algorithms. Follow exactly the corresponding signal flow graphs and keep track of all the intermediate quantities by putting them on the diagrams. 12 Compute the 16-point DFT of the sequence x(n) = cos

π n, 2

0 ≤ n ≤ 15

using the radix-4 decimation-in-time algorithm. 13 Consider the eight-point decimation-in-time (DIT) flow graph in Fig. 1.6. (a) What is the gain of the “signal path” that goes from x(7) to X(2)? (b) How many paths lead from the input to a given output sample? Is this true for every output sample? (c) Compute X(3) using the operations dictated by this flow graph.

569

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

14 Draw the flow graph for the decimation-in-frequency (DIF) SRFFT algorithm for N = 16. What is the number of nontrivial multiplications? 15 Derive the algorithm and draw the N = 8 flow graph for the DIT SRFFT algorithm. Compare your flow graph with the DIF radix-2 FFT flow graph shown in Fig. 1.11. 16 Show that the product of two complex numbers (a+j b) and (c+j d) can be performed with three real multiplications and five additions using the algorithm xR = (a − b)d + (c − d)a xI = (a − b)d + (c + d)b where x = xR + j xI = (a + j b)(c + j d) 17 Explain how the DFT can be used to compute N equispaced samples of the ztransform of an N -point sequence, on a circle of radius r . 18 A real-valued N -point sequence x(n) is called DFT bandlimited if its DFT X(k) = 0 for k0 ≤ k ≤ N − k0 . We insert (L − 1)N zeros in the middle of X(k) to obtain the following LN -point DFT:  

X (k) =

0 ≤ k ≤ k0 − 1 k0 ≤ k ≤ LN − k0 LN − k0 + 1 ≤ k ≤ LN − 1

X(k), 0, X(k + N − LN ),

Show that Lx  (Ln) = x(n),

0≤n≤N −1

where DF T

x  (n) ←→ X  (k) LN

Explain the meaning of this type of processing by working out an example with N = 4, L = 1, and X(k) = {1, 0, 0, 1}. 19 Let X(k) be the N -point DFT of the sequence x(n), 0 ≤ n ≤ N − 1. What is the N -point DFT of the sequence s(n) = X(n), 0 ≤ n ≤ N − 1? 20 Let X(k) be the N -point DFT of the sequence x(n), 0 ≤ n ≤ N − 1. We define a 2N -point sequence y(n) as  n x , y(n) = 2 0,

21

n even n odd

Express the 2N -point DFT of y(n) in terms of X(k). (a) Determine W (z) of the Hanning window  the z-transform  2π n w(n) = 1 − cos N−1 /2. (b) Determine a formula to compute the N -point DFT Xw (k) of the signal xw (n) = w(n)x(n), 0 ≤ n ≤ N − 1, from the N -point DFT X(k) of the signal x(n).

570

Efficient Computation of the DFT: Fast Fourier Transform Algorithms

22 Create a DFT coefficient table that uses only N/4 memory locations to store the first quadrant of the sine sequence (assume N even). 23 Determine the computational burden of the algorithm given by (2.12) and compare it with the computational burden required in the 2N -point DFT of g(n). Assume that the FFT algorithm is a radix-2 algorithm. 24 Consider an IIR system described by the difference equation y(n) = −

N  k=1

ak y(n − k) +

M 

bk x(n − k)

k=0

 Describe a procedure that computes the frequency response H

2π k , k = 0, 1, . . . , N

N − 1 using the FFT algorithm (N = 2ν ). 25 Develop a radix-3 decimation-in-time FFT algorithm for N = 3ν and draw the corresponding flow graph for N = 9. What is the number of required complex multiplications? Can the operations be performed in place? 26 Repeat Problem 25 for the DIF case. 27 FFT input and output pruning In many applications we wish to compute only a few points M of the N -point DFT of a finite-duration sequence of length L (i.e., M 0

(4.21)

On the other hand, for a negative number in two’s-complement representation, the error is 0 ≤ et x < 2E 2−b and hence

0 ≤ et < 2−b+1 ,

x> b, so that we can neglect the factor of 2−bu in the formulas given below. Under these conditions, the probability density functions for the round-off and truncation errors in the two fixed-point representations are illustrated in Fig. 4.4. We note that

625

Implementation of Discrete-Time Systems

Quantizer Q(x)

x

x+ε

(a)

Figure 4.3

Additive noise model for the nonlinear quantization process: (a) actual system; (b) model for quantization.

x

x+ε

+

ε (b)

in the case of truncation of the two’s-complement representation of the number, the average value of the error has a bias of 2−b /2, whereas in all other cases just illustrated, the error has an average value of zero. We shall use this statistical characterization of the quantization errors in our treatment of such errors in digital filtering and in the computation of the DFT for fixed-point implementation.

p(Er) 1 ∆ ∆ = 2 −b −

∆ 2

0

∆ 2

Er

(a) p(Et) 1 2∆ ∆ = 2 −b −∆

0



Et

(b) p(Et)

1 ∆

Figure 4.4

Statistical characterization of quantization errors: (a) round-off error; (b) truncation error for sign-magnitude; (c) truncation error for two’s complement.

626

∆ = 2 −b −∆

0

(c)



Et

Implementation of Discrete-Time Systems

5

Quantization of Filter Coefficients In the realization of FIR and IIR filters in hardware or in software on a generalpurpose computer, the accuracy with which filter coefficients can be specified is limited by the word length of the computer or the length of the register provided to store the coefficients. Since the coefficients used in implementing a given filter are not exact, the poles and zeros of the system function will, in general, be different from the desired poles and zeros. Consequently, we obtain a filter having a frequency response that is different from the frequency response of the filter with unquantized coefficients. In Section 5.1, we demonstrate that the sensitivity of the filter frequency response characteristics to quantization of the filter coefficients is minimized by realizing a filter having a large number of poles and zeros as an interconnection of second-order filter sections. This leads us to the parallel-form and cascade-form realizations in which the basic building blocks are second-order filter sections.

5.1

Analysis of Sensitivity to Quantization of Filter Coefficients

To illustrate the effect of quantization of the filter coefficients in a direct-form realization of an IIR filter, let us consider a general IIR filter with system function M 

bk z−k

k=0 N 

H (z) =

1+

(5.1) ak z

−k

k=1

The direct-form realization of the IIR filter with quantized coefficients has the system function M  bk z−k H (z) =

k=0 N 

1+

(5.2) a k z−k

k=1

where the quantized coefficients {bk }and {a k } can be related to the unquantized coefficients {bk } and {ak } by the relations a k = ak + ak ,

k = 1, 2, . . . , N

bk = bk + bk ,

k = 0, 1, . . . , M

(5.3)

and {ak } and {bk } represent the quantization errors. The denominator of H (z) may be expressed in the form D(z) = 1 +

N  k=0

ak z

−k

=

N 

(1 − pk z−1 )

(5.4)

k=1

627

Implementation of Discrete-Time Systems

where {pk } are the poles of H (z). Similarly, we can express the denominator of H (z) as N  D(z) = (1 − pk z−1 ) (5.5) k=1

where p k = pk +pk , k = 1, 2, . . . , N , and pk is the error or perturbation resulting from the quantization of the filter coefficients. We shall now relate the perturbation pk to the quantization errors in the {ak }. The perturbation error pi can be expressed as pi =

N  ∂pi ak ∂ak

(5.6)

k=1

where ∂pi /∂ak , the partial derivative of pi with respect to ak , represents the incremental change in the pole pi due to a change in the coefficient ak . Thus the total error pi is expressed as a sum of the incremental errors due to changes in each of the coefficients {ak }. The partial derivatives ∂pi /∂ak , k = 1, 2, . . . , N, can be obtained by differentiating D(z) with respect to each of the {ak }. First we have 

∂D(z) ∂ak

Then



 =

z=pi

∂D(z) ∂z



 z=pi

∂pi ∂ak

 (5.7)

(∂D(z)/∂ak )z=pi ∂pi = ∂ak (∂D(z)/∂z)z=pi

(5.8)

The numerator of (5.8) is 

∂D(z) ∂ak



= −z−k |z=pi = −pi−k

(5.9)

z=pi

The denominator of (5.8) is 

∂D(z) ∂z



 = z=pi

 N ∂  −1 (1 − pl z ) ∂z l=1

=

=

   N N  pk   z2   k=1

(1 − pl z−1 )

l=1 l=i

N 1  (pi − pl ) piN l=1 l=i

628

z=pi

      

z=pi

(5.10)

Implementation of Discrete-Time Systems

Therefore, (5.8) can be expressed as −piN−k ∂pi = N ∂ak  (pi − pl )

(5.11)

l=1 l=i

Substitution of the result in (5.11) into (5.6) yields the total perturbation error pi in the form N  piN−k ak (5.12) pi = − N  k=1 (pi − pl ) l=1 l=i

This expression provides a measure of the sensitivity of the i th pole to changes in the coefficients {ak }. An analogous result can be obtained for the sensitivity of the zeros to errors in the parameters {bk }. The terms (pi − pl ) in the denominator of (5.12) represent vectors in the zplane from the poles {pl } to the pole pi . If the poles are tightly clustered as they are in a narrowband filter, as illustrated in Fig. 5.1, the lengths |p i − pl | are small for the poles in the vicinity of pi . These small lengths will contribute to large errors and hence a large perturbation error pi results. The error pi can be minimized by maximizing the lengths |pi − pl |. This can be accomplished by realizing the high-order filter with either single-pole or doublepole filter sections. In general, however, single-pole (and single-zero) filter sections have complex-valued poles and require complex-valued arithmetic operations for their realization. This problem can be avoided by combining complex-valued poles (and zeros) to form second-order filter sections. Since the complex-valued poles are Im(z)

Unit circle

Re(z)

Figure 5.1

Pole positions for a bandpass IIR filter.

629

Implementation of Discrete-Time Systems

x(n)

+

y(n)

z −1

+

2r cos θ

z −1

Figure 5.2

Realization of a two-pole IIR filter.

− r2

usually sufficiently far apart, the perturbation errors {pi } are minimized. As a consequence, the resulting filter with quantized coefficients more closely approximates the frequency response characteristics of the filter with unquantized coefficients. It is interesting to note that even in the case of a two-pole filter section, the structure used to realize the filter section plays an important role in the errors caused by coefficient quantization. To be specific, let us consider a two-pole filter with system function 1 H (z) = (5.13) 1 − (2r cos θ)z−1 + r 2 z−2 This filter has poles at z = re±j θ . When realized as shown in Fig. 5.2, it has two coefficients, a1 = 2r cos θ and a2 = −r 2 . With infinite precision it is possible to achieve an infinite number of pole positions. Clearly, with finite precision (i.e., quantized coefficients a1 and a2 ), the possible pole positions are also finite. In fact, when b bits are used to represent the magnitudes of a1 and a2 , there are at most (2b −1)2 possible positions for the poles in each quandrant, excluding the case a1 = 0 and a2 = 0. For example, suppose that b = 4. Then there are 15 possible nonzero values for a1 . There are also 15 possible values for r 2 . We illustrate these possible values in Fig. 5.3 for the first quandrant of the z-plane only. There are 169 possible pole positions in this case. The nonuniformity in their positions is due to the fact that we are quantizing r 2 , whereas the pole positions lie on a circular arc of radius r . Of particular significance is the sparse set of poles for values of θ near zero and, due to symmetry, near θ = π . This situation would be highly unfavorable for lowpass filters and highpass filters which normally have poles clustered near θ = 0 and θ = π . An alternative realization of the two-pole filter is the coupled-form realization illustrated in Fig. 5.4. The two coupled equations are y1 (n) = x(n) + r cos θ y1 (n − 1) − r sin θ y(n − 1) y(n) = r sin θ y1 (n − 1) + r cos θ y(n − 1)

(5.14)

By transforming these two equations into the z-domain, it is a simple matter to show that Y (z) (r sin θ)z−1 = H (z) = (5.15) X(z) 1 − (2r cos θ)z−1 + r 2 z−2

630

Implementation of Discrete-Time Systems

✶ ✶ ✶ ✶ ✶

✶ ✶ ✶

✶ ✶ ✶

✶ ✶

✶ ✶













✶ ✶ ✶

































✶ ✶ ✶ ✶ ✶ ✶ ✶ ✶ ✶

✶ ✶ ✶ ✶ ✶ ✶







✶ ✶























































































✶ ✶



✶ ✶



✶ ✶











✶ ✶ ✶ ✶ ✶

✶ ✶



✶ ✶







✶ ✶





✶ ✶









✶ ✶

✶ ✶











Possible pole positions for two-pole IIR filter realization in Fig. 5.2.





✶ ✶





✶ ✶



Figure 5.3









✶ ✶ ✶

✶ ✶









In the coupled form we observe that there are also two coefficients, α1 = r sin θ and α2 = r cos θ . Since they are both linear in r , the possible pole positions are now equally spaced points on a rectangular grid, as shown in Fig. 5.5. As a consequence, the pole positions are now uniformly distributed inside the unit circle, which is a more desirable situation than the previous realization, especially for lowpass filters. (There are 198 possible pole positions in this case.) However, the price that we pay for this uniform distribution of pole positions is an increase in computations. The coupled-form realization requires four multiplications per output point, whereas the realization in Fig. 5.2 requires only two multiplications per output point. x(n)

+

y1(n)

+

r cos θ

z −1 y1(n − 1) r sin θ

− r sin θ

y(n)

+ r cos θ

Figure 5.4

Coupled-form realization of a two-pole IIR filter.

z −1 y(n − 1)

631

Implementation of Discrete-Time Systems

Figure 5.5

Possible pole positions for the coupled-form two-pole filter in Fig. 5.4.













































































































































































































































































































































































































































Since there are various ways in which one can realize a second-order filter section, there are obviously many possibilities for different pole locations with quantized coefficients. Ideally, we should select a structure that provides us with a dense set of points in the regions where the poles lie. Unfortunately, however, there is no simple and systematic method for determining the filter realization that yields this desired result. Given that a higher-order IIR filter should be implemented as a combination of second-order sections, we still must decide whether to employ a parallel configuration or a cascade configuration. In other words, we must decide between the realization H (z) =

K  bk0 + bk1 z−1 + bk2 z−2 k=1

and the realization H (z) =

K  k=1

1 + ak1 z−1 + ak2 z−2

ck0 + ck1 z−1 1 + ak1 z−1 + ak2 z−2

(5.16)

(5.17)

If the IIR filter has zeros on the unit circle, as is generally the case with elliptic and Chebyshev type II filters, each second-order section in the cascade configuration of (5.16) contains a pair of complex-conjugate zeros. The coefficients {bk } directly determine the location of these zeros. If the {bk } are quantized, the sensitivity of the system response to the quantization errors is easily and directly controlled by allocating a sufficiently large number of bits to the representation of the {bki }. In fact, we can easily evaluate the perturbation effect resulting from quantizing the

632

Implementation of Discrete-Time Systems

coefficients {bki } to some specified precision. Thus we have direct control of both the poles and the zeros that result from the quantization process. On the other hand, the parallel realization of H (z) provides direct control of the poles of the system only. The numerator coefficients {ck0 } and {ck1 } do not specify the location of the zeros directly. In fact, the {ck0 } and {ck1 } are obtained by performing a partial-fraction expansion of H (z). Hence they do not directly influence the location of the zeros, but only indirectly through a combination of all the factors of H (z). As a consequence, it is more difficult to determine the effect of quantization errors in the coefficients {cki } on the location of the zeros of the system. It is apparent that quantization of the parameters {cki } is likely to produce a significant perturbation of the zero positions and usually, it is sufficiently large in fixed-point implementations to move the zeros off the unit circle. This is a highly undesirable situation, which can be easily remedied by use of a floating-point representation. In any case the cascade form is more robust in the presence of coefficient quantization and should be the preferred choice in practical applications, especially where a fixed-point representation is employed. EXAMPLE 5.1 Determine the effect of parameter quantization on the frequency response of the seventhorder elliptic filter when it is realized as a cascade of second-order sections.

The coefficients for an elliptic filter are specified for the cascade form to six significant digits. We quantized these coefficients to four and then three significant digits (by rounding) and plotted the magnitude (in decibels) and the phase of

Gain (dB)

Solution.

10 0 −10 −20 −30 −40 −50 −60 −70 −80 −90 −100 0

Unquantized Quantized to 3 and 4 digits

.1

.2

.3

.4

.5

.4

.5

f

180

Figure 5.6

Effect of coefficient quantization of the magnitude and phase response of an N = 7 elliptic filter realized in cascade form.

Phase (degree)

120 60 0 − 60 −120 −180 0

.1

.2 .3 Relative frequency

f

633

Implementation of Discrete-Time Systems

the frequency response. The results are shown in Fig. 5.6 along the frequency response of the filter with unquantized (six significant digits) coefficients. We observe that there is an insignificant degradation due to coefficient quantization for the cascade realization.

EXAMPLE 5.2 Repeat the computation of the frequency response for the elliptic filter considered in Example 5.1 when it is realized in the parallel form with second-order sections. Solution.

The system function for a 7-order elliptic filter is H (z) =

0.2781304 + 0.0054373108z−1 1 − 0.790103z−1 +

−0.3867805 + 0.3322229z−1 1 − 1.517223z−1 + 0.714088z−2

+

0.1277036 − 0.1558696z−1 1 − 1.421773z−1 + 0.861895z−2

+

−0.015824186 + 0.38377356z−1 1 − 1.387447z−1 + 0.962242z−2

The frequency response of this filter with coefficients quantized to four digits is shown in Fig. 5.7(a). When this result is compared with the frequency response in Fig. 5.6, we observe that the zeros in the parallel realization have been perturbed sufficiently so that the nulls in the magnitude response are now at −80, −85, and −92 dB. The phase response has also been perturbed by a small amount. When the coefficients are quantized to three significant digits, the frequency response characteristic deteriorates significantly, in both magnitude and phase, as illustrated in Fig. 5.7(b). It is apparent from the magnitude response that the zeros are no longer on the unit circle as a result of the quantization of the coefficients. This result clearly illustrates the sensitivity of the zeros to quantization of the coefficients in the parallel form. When compared with the results of Example 5.1, it is also apparent that the cascade form is definitely more robust to parameter quantization than the parallel form.

5.2

Quantization of Coefficients in FIR Filters

As indicated in the preceding section, the sensitivity analysis performed on the poles of a system also applies directly to the zeros of the IIR filters. Consequently, an expression analogous to (5.12) can be obtained for the zeros of an FIR filter. In effect, we should generally realize FIR filters with a large number of zeros as a cascade of second-order and first-order filter sections to minimize the sensitivity to coefficient quantization. Of particular interest in practice is the realization of linear-phase FIR filters. The direct-form realizations shown in Figs. 2.1 and 2.2 maintain the linear-phase property even when the coefficients are quantized. This follows easily from the observation that the system function of a linear-phase FIR filter satisfies the property H (z) = ±z−(M−1) H (z−1 )

634

Gain (dB)

Implementation of Discrete-Time Systems

10 0 −10 −20 −30 −40 −50 −60 −70 −80 −90 −100 0

.1

.2

.3

.4

.5

.4

.5

.4

.5

.4

.5

f

180

Phase (degree)

120 60 0 −60

Gain (dB)

−120 −180 0

.1

10 0 −10 −20 −30 −40 −50 −60 −70 −80 −90 −100 0

.1

.2 .3 Relative frequency (a) Quantization to 4 digits

.2

.3

f

f

180

Effect of coefficient quantization of the magnitude and phase response of an N = 7 elliptic filter realized in parallel form: (a) quantization to four digits; (b) quantization to three digits.

Phase (degree)

120

Figure 5.7

60 0 −60 −120 −180 0

.1

.2 .3 Relative frequency (b) Quantization to 3 digits

f

635

Implementation of Discrete-Time Systems

independent of whether the coefficients are quantized or unquantized. Consequently, coefficient quantization does not affect the phase characteristic of the FIR filter, but affects only the magnitude. As a result, coefficient quantization effects are not as severe on a linear-phase FIR filter, since the only effect is in the magnitude. EXAMPLE 5.3 Determine the effect of parameter quantization on the frequency response of an M = 32 linear-phase FIR bandpass filter. The filter is realized in the direct form. Solution. The frequency response of a linear-phase FIR bandpass filter with unquantized coefficients is illustrated in Fig. 5.8(a). When the coefficients are quantized to four significant digits, the effect on the frequency response is insignificant. However, when the coefficients are quantized to three significant digits, the sidelobes increase by several decibels, as illustrated in Fig. 5.8(b). This result indicates that we should use a minimum of 10 bits to represent the coefficients of this FIR filter and, preferably, 12 to 14 bits, if possible.

From this example we learn that a minimum of 10 bits is required to represent the coefficients in a direct-form realization of an FIR filter of moderate length. As the filter length increases, the number of bits per coefficient must be increased to maintain the same error in the frequency response characteristic of the filter. For example, suppose that each filter coefficient is rounded to (b + 1) bits. Then the maximum error in a coefficient value is bounded as

Gain (dB)

−2−(b+1) < eh (n) < 2−(b+1) 10 0 −10 −20 −30 −40 −50 −60 −70 −80 −90 −100 0

.1

.2 .3 Relative frequency

.4

.5

.4

.5

f

Figure 5.8

Effect of coefficient quantization of the magnitude of an M = 32 linear-phase FIR filter realized in direct form: (a) no quantization; (b) quantization to three digits.

636

Gain (dB)

(a) No quantization 10 0 −10 −20 −30 −40 −50 −60 −70 −80 −90 −100 0

.1

.2 .3 Relative frequency (b) Quantization to 3 digits

f

Implementation of Discrete-Time Systems

Since the quantized values may be represented as h(n) = h(n) + eh (n), the error in the frequency response is EM (ω) =

M−1 

eh (n)e−j ωn

n=0

Since eh (n) is zero mean, it follows that EM (ω) is also zero mean. Assuming that the coefficient error sequence eh (n), 0 ≤ n ≤ M − 1, is uncorrelated, the variance of the error EM (ω) in the frequency response is just the sum of the variances of the M terms. Thus we have 2−2(b+1) 2−2(b+2) σE2 = M= M 12 3 Here we note that the variance of the error in H (ω) increases linearly with M . Hence the standard deviation of the error in H (ω) is σE =

2−(b+2) √ M √ 3

Consequently, for every factor-of-4 increase in M , the precision in the filter coefficients must be increased by one additional bit to maintain the standard deviation fixed. This result, taken together with the results of Example 5.3, implies that the frequency error remains tolerable for filter lengths up to 256, provided that filter coefficients are represented by 12 to 13 bits. If the word length of the digital signal processor is less than 12 bits or if the filter length exceeds 256, the filter should be implemented as a cascade of smaller length filters to reduce the precision requirements. In a cascade realization of the form H (z) = G

K 

Hk (z)

(5.18)

k=1

where the second-order sections are given as Hk (z) = 1 + bk1 z−1 + bk2 z−2

(5.19)

the coefficients of complex-valued zeros are expressed as bk1 = −2rk cos θk and bk2 = rk2 . Quantization of bk1 and bk2 results in zero locations as shown in Fig. 5.3, except that the grid extends to points outside the unit circle. A problem may arise, in this case, in maintaining the linear-phase property, because the quantized pair of zeros at z = (1/rk )e±j θk may not be the mirror image of the quantized zeros at z = rk e±j θk . This problem can be avoided by rearranging the factors corresponding to the mirror-image zero. That is, we can write the mirrorimage factor as   2 1 −2 1 −1 (5.20) 1 − cos θk z + 2 z = 2 (rk2 − 2rk cos θk z−1 + z−2 ) rk rk rk

637

Implementation of Discrete-Time Systems

The factors {1/rk2 } can be combined with the overall gain factor G, or they can be distributed in each of the second-order filters. The factor in (5.20) contains exactly the same parameters as the factor (1 − 2rk cos θk z−1 + rk2 z−2 ), and consequently, the zeros now occur in mirror-image pairs even when the parameters are quantized. In this brief treatment we have given the reader an introduction to the problems of coefficient quantization in IIR and FIR filters. We have demonstrated that a highorder filter should be reduced to a cascade (for FIR or IIR filters) or a parallel (for IIR filters) realization to minimize the effects of quantization errors in the coefficients. This is especially important in fixed-point realizations in which the coefficients are represented by a relatively small number of bits.

6

Round-Off Effects in Digital Filters In Section 4 we characterized the quantization errors that occur in arithmetic operations performed in a digital filter. The presence of one or more quantizers in the realization of a digital filter results in a nonlinear device with characteristics that may be significantly different from the ideal linear filter. For example, a recursive digital filter may exhibit undesirable oscillations in its output, as shown in the following section, even in the absence of an input signal. As a result of the finite-precision arithmetic operations performed in the digital filter, some registers may overflow if the input signal level becomes large. Overflow represents another form of undesirable nonlinear distortion on the desired signal at the output of the filter. Consequently, special care must be exercised to scale the input signal properly, either to prevent overflow completely or, at least, to minimize its rate of occurrence. The nonlinear effects due to finite-precision arithmetic make it extremely difficult to precisely analyze the performance of a digital filter. To perform an analysis of quantization effects, we adopt a statistical characterization of quantization errors which, in effect, results in a linear model for the filter. Thus we are able to quantify the effects of quantization errors in the implementation of digital filters. Our treatment is limited to fixed-point realizations where quantization effects are very important.

6.1

Limit-Cycle Oscillations in Recursive Systems

In the realization of a digital filter, either in digital hardware or in software on a digital computer, the quantization inherent in the finite-precision arithmetic operations renders the system nonlinear. In recursive systems, the nonlinearities due to the finite-precision arithmetic operations often cause periodic oscillations to occur in the output, even when the input sequence is zero or some nonzero constant value. Such oscillations in recursive systems are called limit cycles and are directly attributable to round-off errors in multiplication and overflow errors in addition. To illustrate the characteristics of a limit-cycle oscillation, let us consider a singlepole system described by the linear difference equation y(n) = ay(n − 1) + x(n)

638

(6.1)

Implementation of Discrete-Time Systems

x(n)

y(n)

+

z −1

Figure 6.1

Ideal single-pole recursive system.

a

where the pole is at z = a . The ideal system is realized as shown in Fig. 6.1. On the other hand, the actual system, which is described by the nonlinear difference equation v(n) = Q[av(n − 1)] + x(n)

(6.2)

is realized as shown in Fig. 6.2. Suppose that the actual system in Fig. 6.2 is implemented with fixed-point arithmetic based on four bits for the magnitude plus a sign bit. The quantization that takes place after multiplication is assumed to round the resulting product upward. In Table 2 we list the response of the actual system for four different locations of the pole z = a , and an input x(n) = βδ(n), where β = 15/16, which has the binary representation 0.1111. Ideally, the response of the system should decay toward zero exponentially [i.e., y(n) = a n → 0 as n → ∞]. In the actual system, however, the response v(n) reaches a steady-state periodic output sequence with a period that depends on the value of the pole. When the pole is positive, the oscillations 1 occur with a period Np = 1, so that the output reaches a constant value of 16 for 1 1 3 a = 2 and 8 for a = 4 . On the other hand, when the pole is negative, the output 1 for a = − 21 and ± 18 sequence oscillates between positive and negative values (± 16 for a = − 43 ). Hence the period is Np = 2. These limit cycles occur as a result of the quantization effects in multiplications. When the input sequence x(n) to the filter becomes zero, the output of the filter then, after a number of iterations, enters into the limit cycle. The output remains in the limit cycle until another input of sufficient size is applied that drives the system out of the limit cycle. Similarly, zero-input limit cycles occur from nonzero initial conditions with the input x(n) = 0. The amplitudes of the output during a limit cycle are confined to a range of values that is called the dead band of the filter. It is interesting to note that when the response of the single-pole filter is in the limit cycle, the actual nonlinear system operates as an equivalent linear system with x(n)

ν(n)

+

z −1

Q[ ]

Figure 6.2

Actual nonlinear system.

a

639

Implementation of Discrete-Time Systems

TABLE 2 Limit Cycles for Lowpass Single-Pole Filter

n 0 1 2 3 4 5 6 7 8

a = 0.1000 = 21  15  0.1111  16  8 0.1000 16 4 0.0100  16  2 0.0010  16  1 0.0001 16 1 0.0001  16  1 0.0001 16 1 0.0001  16  1 0.0001 16

a = 1.1000 = − 21  15  0.1111  16  8 1.1000 − 16 4 0.0100  16  2 1.0010 − 16 1 0.0001  16  1 1.0001 − 16 1 0.0001  16  1 1.0001 − 16 1 0.0001 16

a = 0.1100 = 43  11  0.1011  16  8 0.1000 16 6 0.0110  16  5 0.0101  16  4 0.0100 16 3 0.0011  16  2 0.0010 16 2 0.0010  16  2 0.0010 16

a = 1.1100 = − 43  11  0.1011  16  8 1.1000 − 16 6 0.0110  16  5 1.0101 − 16 4 0.0100  16  3 1.0011 − 16 2 0.0010  16  2 1.0010 − 16 2 0.0010 16

a pole at z = 1 when the pole is positive and z = −1 when the pole is negative. That is,  v(n − 1), a>0 Qr [av(n − 1)] = (6.3) −v(n − 1), a 0 −ω  2  (ω) = M −1   + π, if Hr (ω) < 0  −ω 2

(2.8)

(2.9)

(2.10)

When h(n) = −h(M − 1 − n) the unit sample response is antisymmetric. For M odd, the center point of the antisymmetric h(n) is n = (M − 1)/2. Consequently,  M −1 =0 h 2 However, if M is even, each term in h(n) has a matching term of opposite sign.

678

Design of Digital Filters

It is straightforward to show that the frequency response of an FIR filter with an antisymmetric unit sample response can be expressed as H (ω) = Hr (ω)ej [−ω(M−1)/2+π/2]

(2.11)

where 



(M−3)/2

Hr (ω) = 2

h(n) sin ω

n=0





(M/2)−1

Hr (ω) = 2

h(n) sin ω

n=0

M −1 −n , 2

M odd

(2.12)

M −1 −n , 2

M even

(2.13)

The phase characteristic of the filter for both M odd and M even is   π M −1   , if Hr (ω) > 0  −ω 2  2 (ω) = 3π M −1    −ω , if Hr (ω) < 0 2 2

(2.14)

These general frequency response formulas can be used to design linear-phase FIR filters with symmetric and antisymmetric unit sample responses. We note that, for a symmetric h(n), the number of filter coefficients that specify the frequency response is (M + 1)/2 when M is odd or M/2 when M is even. On the other hand, if the unit sample response is antisymmetric,  M −1 h =0 2 so that there are (M − 1)/2 filter coefficients when M is odd and M/2 coefficients when M is even to be specified. The choice of a symmetric or antisymmetric unit sample response depends on the application. As we shall see later, a symmetric unit sample response is suitable for some applications, while an antisymmetric unit sample response is more suitable for other applications. For example, if h(n) = −h(M − 1 − n) and M is odd, (2.12) implies that Hr (0) = 0 and Hr (π ) = 0. Consequently, (2.12) is not suitable as either a lowpass filter or a highpass filter. Similarly, the antisymmetric unit sample response with M even also results in Hr (0) = 0, as can be easily verified from (2.13). Consequently, we would not use the antisymmetric condition in the design of a lowpass linear-phase FIR filter. On the other hand, the symmetry condition h(n) = h(M − 1 − n) yields a linear-phase FIR filter with a nonzero response at ω = 0, if desired, that is,  (M−3)/2  M −1 Hr (0) = h h(n), M odd +2 (2.15) 2 n=0



(M/2)−1

Hr (0) = 2

h(n),

M even

(2.16)

n=0

679

Design of Digital Filters

In summary, the problem of FIR filter design is simply to determine the M coefficients h(n), n = 0, 1, . . . , M − 1, from a specification of the desired frequency response Hd (ω) of the FIR filter. The important parameters in the specification of Hd (ω) are given in Fig. 1.2. In the following subsections we describe design methods based on specification of Hd (ω).

2.2

Design of Linear-Phase FIR Filters Using Windows

In this method we begin with the desired frequency response specification Hd (ω) and determine the corresponding unit sample response hd (n). Indeed, hd (n) is related to Hd (ω) by the Fourier transform relation Hd (ω) =

∞ 

hd (n)e−j ωn

(2.17)

Hd (ω)ej ωn dω

(2.18)

n=0

where hd (n) =

1 2π



π

−π

Thus, given Hd (ω), we can determine the unit sample response hd (n) by evaluating the integral in (2.17). In general, the unit sample response hd (n) obtained from (2.17) is infinite in duration and must be truncated at some point, say at n = M − 1, to yield an FIR filter of length M . Truncation of hd (n) to a length M − 1 is equivalent to multiplying hd (n) by a “rectangular window,” defined as  w(n) =

n = 0, 1, . . . , M − 1 otherwise

1, 0,

(2.19)

Thus the unit sample response of the FIR filter becomes h(n) = hd (n)w(n)  hd (n), = 0,

n = 0, 1, . . . , M − 1 otherwise

(2.20)

It is instructive to consider the effect of the window function on the desired frequency response Hd (ω). Recall that multiplication of the window function w(n) with hd (n) is equivalent to convolution of Hd (ω) with W (ω), where W (ω) is the frequency-domain representation (Fourier transform) of the window function, that is, W (ω) =

M−1  n=0

680

w(n)e−j ωn

(2.21)

Design of Digital Filters

Thus the convolution of Hd (ω) with W (ω) yields the frequency response of the (truncated) FIR filter. That is, H (ω) =

1 2π



π

−π

Hd (ν)W (ω − ν)dν

(2.22)

The Fourier transform of the rectangular window is W (ω) =

M−1 

e−j ωn

n=0

(2.23)

1 − e−j ωM sin(ωM/2) = = e−j ω(M−1)/2 −j ω 1−e sin(ω/2) This window function has a magnitude response |W (ω)| =

| sin(ωM/2)| , | sin(ω/2)|

and a piecewise linear phase   M −1   ,  −ω  2 (ω) = M −1   + π,  −ω 2

π ≤ω≤π

(2.24)

when sin(ωM/2) ≥ 0 (2.25) when sin(ωM/2) < 0

The magnitude response of the window function is illustrated in Fig. 2.2 for M = 31 and 61. The width of the main lobe [width is measured to the first zero of W (ω)] is 4π/M . Hence, as M increases, the main lobe becomes narrower. However, the sidelobes of |W (ω)| are relatively high and remain unaffected by an increase in M . In fact, even though the width of each sidelobe decreases with an increase in M , the height of each sidelobe increases with an increase in M in such a manner that the area under each sidelobe remains invariant to changes in M . This characteristic behavior is not evident from observation of Fig. 2.2 because W (ω) has been normalized by M such that the normalized peak values of the sidelobes remain invariant to an increase in M .

Figure 2.2

Frequency response for rectangular window of lengths (a) M = 31, (b) M = 61.

681

Design of Digital Filters

The characteristics of the rectangular window play a significant role in determining the resulting frequency response of the FIR filter obtained by truncating hd (n) to length M . Specifically, the convolution of Hd (ω) with W (ω) has the effect of smoothing Hd (ω). As M is increased, W (ω) becomes narrower, and the smoothing provided by W (ω) is reduced. On the other hand, the large sidelobes of W (ω) result in some undesirable ringing effects in the FIR filter frequency response H (ω), and also in relatively larger sidelobes in H (ω). These undesirable effects are best alleviated by the use of windows that do not contain abrupt discontinuities in their time-domain characteristics, and have correspondingly low sidelobes in their frequency-domain characteristics. Table 1 lists several window functions that possess desirable frequency response characteristics. Figure 2.3 illustrates the time-domain characteristics of the windows. The frequency response characteristics of the Hanning, Hamming, and Blackman windows are illustrated in Figs. 2.4 through 2.6. All of these window TABLE 1 Window Functions for FIR Filter Design

Name of

Time-domain sequence,

window

h(n), 0 ≤ n ≤ M − 1    M − 1   2 n − 2  1− M −1

Bartlett (triangular) Blackman Hamming Hanning

Kaiser

Lanczos

Tukey

682

0.42 − 0.5 cos

4πn 2πn + 0.08 cos M −1 M −1

2πn M −1  2πn 1 1 − cos 2 M −1

0.54 − 0.46 cos

   2  2  M − 1 M − 1  − n− I0 α 2 2    M −1 I0 α 2   L    M −1     (M − 1) sin 2π n −   2    , L>0   M −1 M −1     2π n − 2 2    M −1 M − 1  ≤α , 0 1, then σ > 0. Consequently, the LHP in s maps into the inside of the unit circle in the z-plane and the RHP in s maps into the outside of the unit circle. When r = 1, then σ = 0 and =

2 sin ω T 1 + cos ω

(3.43)

2 ω = tan T 2 or, equivalently, ω = 2 tan−1

T 2

(3.44)

The relationship in (3.44) between the frequency variables in the two domains is illustrated in Fig. 3.8. We observe that the entire range in is mapped only once into the range −π ≤ ω ≤ π . However, the mapping is highly nonlinear. We observe a frequency compression or frequency warping, as it is usually called, due to the nonlinearity of the arctangent function.

730

Design of Digital Filters

ω ω = 2 tan−1 ΩT 2

π π 2

−10

−5

5

0

10

15

ΩT

−π 2 −π

Figure 3.8 Mapping between the frequency variables ω and

resulting from the bilinear transformation.

It is also interesting to note that the bilinear transformation maps the point s = ∞ into the point z = −1. Consequently, the single-pole lowpass filter in (3.33), which has a zero at s = ∞, results in a digital filter that has a zero at z = −1. EXAMPLE 3.4 Convert the analog filter with system function Ha (s) =

s + 0.1 (s + 0.1)2 + 16

into a digital IIR filter by means of the bilinear transformation. The digital filter is to have a resonant frequency of ωr = π/2. Solution. First, we note that the analog filter has a resonant frequency r = 4. This frequency is to be mapped into ωr = π/2 by selecting the value of the parameter T . From the relationship in (3.43), we must select T = 21 in order to have ωr = π/2. Thus the desired mapping is  1 − z−1 s=4 1 + z−1 The resulting digital filter has the system function H (z) =

0.125 + 0.0061z−1 − 0.1189z−1 1 + 0.0006z−1 + 0.9512z−2

We note that the coefficient of the z−1 term in the denominator of H (z) is extremely small and can be approximated by zero. Thus we have the system function H (z) =

0.125 + 0.0061z−1 − 0.1189z−2 1 + 0.9512 z −2

731

Design of Digital Filters

This filter has poles at

p1,2 = 0.987e±j π/2

and zeros at z1,2 = −1, 0.95 Therefore, we have succeeded in designing a two-pole filter that resonates near ω = π/2.

In this example the parameter T was selected to map the resonant frequency of the analog filter into the desired resonant frequency of the digital filter. Usually, the design of the digital filter begins with specifications in the digital domain, which involve the frequency variable ω. These specifications in frequency are converted to the analog domain by means of the relation in (3.43). The analog filter is then designed that meets these specifications and converted to a digital filter by means of the bilinear transformation in (3.40). In this procedure, the parameter T is transparent and may be set to any arbitrary value (e.g., T = 1). The following example illustrates this point. EXAMPLE 3.5 Design a single-pole lowpass digital filter with a 3-dB bandwidth of 0.2π , using the bilinear transformation applied to the analog filter H (s) =

c s + c

where c is the 3-dB bandwidth of the analog filter. Solution. The digital filter is specified to have its −3-dB gain at ωc = 0.2π . In the frequency domain of the analog filter ωc = 0.2π corresponds to 2 tan 0.1π T

c =

0.65 T

=

Thus the analog filter has the system function H (s) =

0.65/T s + 0.65/T

This represents our filter design in the analog domain. Now, we apply the bilinear transformation given by (3.40) to convert the analog filter into the desired digital filter. Thus we obtain H (z) =

0.245(1 + z−1 ) 1 − 0.509z−1

where the parameter T has been divided out. The frequency response of the digital filter is H (ω) =

0.245(1 + e−j ω ) 1 − 0.509e−j ω

At ω = 0, H (0) = 1, and at ω = 0.2π , we have |H (0.2π)| = 0.707, which is the desired response.

732

Design of Digital Filters

3.4

Characteristics of Commonly Used Analog Filters

As we have seen from our prior discussion, IIR digital filters can easily be obtained by beginning with an analog filter and then using a mapping to transform the s plane into the z-plane. Thus the design of a digital filter is reduced to designing an appropriate analog filter and then performing the conversion from H (s) to H (z), in such a way so as to preserve, as much as possible, the desired characteristics of the analog filter. Analog filter design is a well-developed field and many books have been written on the subject. In this section we briefly describe the important characteristics of commonly used analog filters and introduce the relevant filter parameters. Our discussion is limited to lowpass filters. Subsequently, we describe several frequency transformations that convert a lowpass prototype filter into either a bandpass, highpass, or band-elimination filter. Butterworth filters. Lowpass Butterworth filters are all-pole filters characterized by the magnitude-squared frequency response |H ( )|2 =

1 1 = 1 + ( / c )2N 1 + 2 ( / p )2N

(3.45)

where N is the order of the filter, c is its −3-dB frequency (usually called the cutoff frequency), p is the passband edge frequency, and 1/(1 + 2 ) is the band-edge value of |H ( )|2 . Since H (s)H (−s) evaluated at s = j is simply equal to |H ( )|2 , it follows that 1 H (s)H (−s) = (3.46) 1 + (−s 2 / 2c )N The poles of H (s)H (−s) occur on a circle of radius c at equally spaced points. From (3.46) we find that −s 2 = (−1)1/N = ej (2k+1)π/N , 2c

k = 0, 1, . . . , N − 1

and hence sk = c ej π/2 ej (2k+1)π/2N ,

k = 0, 1, . . . , N − 1

(3.47)

For example, Fig. 3.9 illustrates the pole positions for N = 4 and N = 5 Butterworth filters. The frequency response characteristics of the class of Butterworth filters are shown in Fig. 3.10 for several values of N . We note that |H ( )|2 is monotonic in both the passband and stopband. The order of the filter required to meet an attenuation δ2 at a specified frequency s is easily determined from (3.45). Thus at = s we have 1 = δ22 1 + 2 ( s / p )2N

733

Design of Digital Filters

π π + 2 8 Poles of H(s)

−π − π 2 8

Poles of H(−s)

N=4 (a)

π π + 2 10 Poles of H(s)

Poles of H(−s)

N=5 (b)

Figure 3.9 Pole positions for Butterworth filters.

and hence

log[(1/δ22 ) − 1] log(δ/ ) (3.48) = 2 log( s / c ) log( s / p ) √ where, by definition, δ2 = 1/ 1 + δ 2 . Thus the Butterworth filter is completely characterized by the parameters N , δ2 , , and the ratio s / p . N=

EXAMPLE 3.6 Determine the order and the poles of a lowpass Butterworth filter that has a −3-dB bandwidth of 500 Hz and an attenuation of 40 dB at 1000 Hz. Solution. The critical frequencies are the −3-dB frequency c and the stopband frequency s , which are c = 1000π s = 2000π

734

Design of Digital Filters

|H(Ω)|2 1.1 1 1.0 1 + ε2 0.9

0.8

0.7 0.6

0.5

0.4

0.3 N=1 0.2 N=3

0.1 N=5 0

N=2

N=4 Ω

Ωe

Figure 3.10 Frequency response of Butterworth filters.

For an attenuation of 40 dB, δ2 = 0.01. Hence from (3.48) we obtain N=

log10 (104 − 1) 2 log10 2

= 6.64 To meet the desired specifications, we select N = 7. The pole positions are sk = 1000πej [π/2+(2k+1)π/14] ,

k = 0, 1, 2, . . . , 6

Chebyshev filters. There are two types of Chebyshev filters. Type I Chebyshev filters are all-pole filters that exhibit equiripple behavior in the passband and a monotonic

735

Design of Digital Filters

characteristic in the stopband. On the other hand, the family of type II Chebyshev filters contains both poles and zeros and exhibits a monotonic behavior in the passband and an equiripple behavior in the stopband. The zeros of this class of filters lie on the imaginary axis in the s -plane. The magnitude squared of the frequency response characteristic of a type I Chebyshev filter is given as |H ( )|2 =

1 1+

2 TN2 ( / p )

(3.49)

where is a parameter of the filter related to the ripple in the passband and TN (x) is the N th-order Chebyshev polynomial defined as  TN (x) =

cos(N cos−1 x), cosh(N cosh−1 x),

|x| ≤ 1 |x| > 1

(3.50)

The Chebyshev polynomials can be generated by the recursive equation TN+1 (x) = 2xTN (x) − TN−1 (x),

N = 1, 2, . . .

(3.51)

where T0 (x) = 1 and T1 (x) = x . From (3.51) we obtain T2 (x) = 2x 2 − 1, T3 (x) = 4x 3 − 3x , and so on. Some of the properties of these polynomials are as follows: 1. |TN (x)| ≤ 1 for all |x| ≤ 1. 2. TN (1) = 1 for all N . 3. All the roots of the polynomial TN (x) occur in the interval −1 ≤ x ≤ 1. The filter parameter is related to the ripple in the passband, as illustrated in Fig. 3.11, for N odd and N even. For N odd, TN (0) = 0 and hence |H (0)|2 = 1. On the other hand, for N even, TN (0) = 1 and hence |H (0)|2 = 1/(1 + 2 ). At the band-edge frequency = p , we have TN (1) = 1, so that √

1 1 + 2

= 1 − δ1

or, equivalently, 2 =

1 −1 (1 − δ1 )2

where δ1 is the value of the passband ripple.

736

(3.52)

Design of Digital Filters

Figure 3.11 Type I Chebyshev filter characteristic.

The poles of a type I Chebyshev filter lie on an ellipse in the s -plane with major axis r1 = p

β2 + 1 2β

(3.53)

r2 = p

β2 − 1 2β

(3.54)

and minor axis

where β is related to according to the equation !√ β=

1 + 2 + 1

"1/N (3.55)

The pole locations are most easily determined for a filter of order N by first locating the poles for an equivalent N th-order Butterworth filter that lie on circles of radius r1 or radius r2 , as illustrated in Fig.3.12. If we denote the angular positions of the poles of the Butterworth filter as φk =

π (2k + 1)π + , 2 2N

k = 0, 1, 2, . . . , N − 1

(3.56)

then the positions of the poles for the Chebyshev filter lie on the ellipse at the coordinates (xk , yk ), k = 0, 1, . . . , N − 1, where xk = r2 cos φk ,

k = 0, 1, . . . , N − 1

yk = r1 sin φk ,

k = 0, 1, . . . , N − 1

(3.57)

737

Design of Digital Filters

r2

r1

Figure 3.12

Determination of the pole locations for a Chebyshev filter.

A type II Chebyshev filter contains zeros as well as poles. The magnitude squared of its frequency response is given as |H ( )|2 =

1 1+

2 [TN2 ( s / p )/TN2 ( s / )]

(3.58)

where TN (x) is, again, the N th-order Chebyshev polynomial and s is the stopband frequency as illustrated in Fig. 3.13. The zeros are located on the imaginary axis at the points s sk = j , k = 0, 1, . . . , N − 1 (3.59) sin φk The poles are located at the points (vk , wk ), where vk = (

wk = (

s xk

,

k = 0, 1, . . . , N − 1

(3.60)

,

k = 0, 1, . . . , N − 1

(3.61)

xk2 + yk2 s yk xk2 + yk2

where {xk } and {yk } are defined in (3.57) with β now related to the ripple in the stopband through the equation  β=

738

1+

(

1 − δ22

δ2

1/N 

(3.62)

Design of Digital Filters

Figure 3.13 Type II Chebyshev filters.

From this description, we observe that the Chebyshev filters are characterized by the parameters N , , δ2 , and the ratio s / p . For a given set of specifications on , δ2 , and s / p , we can determine the order of the filter from the equation  ( ( 2 2 2 1 − δ2 + 1 − δ2 (1 + ) / δ2 log   N= ( 2 log ( s / p ) + ( s / p ) − 1 =

(3.63)

cosh−1 (δ/ ) cosh−1 ( s / p )

√ where, by definition, δ2 = 1/ 1 + δ 2 . EXAMPLE 3.7 Determine the order and the poles of a type I lowpass Chebyshev filter that has a 1-dB ripple in the passband, a cutoff frequency p = 1000π , a stopband frequency of 2000π , and an attenuation of 40 dB or more for ≥ s . Solution.

First, we determine the order of the filter. We have 10 log10 (1 + 2 ) = 1 1 + 2 = 1.259 2 = 0.259 = 0.5088

Also, 20 log10 δ2 = −40 δ2 = 0.01

739

Design of Digital Filters

Hence from (3.63) we obtain N=

log10 196.54 √ log10 (2 + 3)

= 4.0 Thus a type I Chebyshev filter having four poles meets the specifications. The pole positions are determined from the relations in (3.53) through (3.57). First, we compute β , r1 , and r2 . Hence β = 1.429 r1 = 1.06 p r2 = 0.365 p The angles {φk } are φk =

(2k + 1)π π + , 2 8

k = 0, 1, 2, 3

Therefore, the poles are located at x1 + jy1 = −0.1397 p ± j 0.979 p x2 + jy2 = −0.337 p ± j 0.4056 p

The filter specifications in Example 3.7 are very similar to the specifications given in Example 3.6, which involved the design of a Butterworth filter. In that case the number of poles required to meet the specifications was seven. On the other hand, the Chebyshev filter required only four. This result is typical of such comparisons. In general, the Chebyshev filter meets the specifications with fewer poles than the corresponding Butterworth filter. Alternatively, if we compare a Butterworth filter to a Chebyshev filter having the same number of poles and the same passband and stopband specifications, the Chebyshev filter will have a smaller transition bandwidth. For a tabulation of the characteristics of Chebyshev filters and their pole–zero locations, the interested reader is referred to the handbook of Zverev (1967). Elliptic filters. Elliptic (or Cauer) filters exhibit equiripple behavior in both the passband and the stopband, as illustrated in Fig. 3.14 for N odd and N even. This class of filters contains both poles and zeros and is characterized by the magnitude-squared frequency response 1 |H ( )|2 = (3.64) 1 + 2 UN ( / p ) where UN (x) is the Jacobian elliptic function of order N , which has been tabulated by Zverev (1967), and is a parameter related to the passband ripple. The zeros lie on the j -axis.

740

Design of Digital Filters

|H(Ω)|2

|H(Ω)|2 1

1

1 1 + ε2

1 1 + ε2

2

2

δ2 N even



δ2 N odd



Figure 3.14 Magnitude-squared frequency characteristics of elliptic filters.

We recall from our discussion of FIR filters that the most efficient designs occur when we spread the approximation error equally over the passband and the stopband. Elliptic filters accomplish this objective and, as a consequence, are the most efficient from the viewpoint of yielding the smallest-order filter for a given set of specifications. Equivalently, we can say that for a given order and a given set of specifications, an elliptic filter has the smallest transition bandwidth. The filter order required to achieve a given set of specifications in passband ripple δ1 , stopband ripple δ2 , and transition ratio p / s is given as )  K( p / s )K 1 − ( 2 /δ 2 ) N= (3.65)  ( K( /δ)K 1 − ( p / s )2 where K(x) is the complete elliptic integral of the first kind, defined as  π/2 dθ ) (3.66) K(x) = 0 1 − x 2 sin2 θ √ and δ2 = 1/ 1 + δ 2 . Values of this integral have been tabulated in a number of texts [e.g., the books by Jahnke and Emde (1945) and Dwight (1957)]. The passband ripple is 10 log10 (1 + 2 ). We shall not attempt to describe elliptic functions in any detail because such a discussion would take us too far afield. Suffice to say that computer programs are available for designing elliptic filters from the frequency specifications indicated above. In view of the optimality of elliptic filters, the reader may question the reason for considering the class of Butterworth or the class of Chebyshev filters in practical applications. One important reason that these other types of filters might be preferable in some applications is that they possess better phase response characteristics. The phase response of elliptic filters is more nonlinear in the passband than a comparable Butterworth filter or a Chebyshev filter, especially near the band edge.

741

Design of Digital Filters

Bessel filters. Bessel filters are a class of all-pole filters that are characterized by the system function 1 H (s) = (3.67) BN (s) where BN (s) is the N th-order Bessel polynomial. These polynomials can be expressed in the form N  (3.68) ak s k BN (s) = k=0

where the coefficients {ak } are given as ak =

(2N − k)! , − k)!

2N−k k!(N

k = 0, 1, . . . , N

(3.69)

Alternatively, the Bessel polynomials may be generated recursively from the relation BN (s) = (2N − 1)BN−1 (s) + s 2 BN−2 (s)

(3.70)

with B0 (s) = 1 and B1 (s) = s + 1 as initial conditions. An important characteristic of Bessel filters is the linear-phase response over the passband of the filter. For example, Fig. 3.15 shows a comparison of the magnitude Magnitude 1.0 0.8 0.6 Bessel 0.4 Butterworth

0.2 0

N=4 0

1

2

3

4

5

1

2

3

4

5



Phase 180 135 90 45 0

Figure 3.15

Magnitude and phase responses of Bessel and Butterworth filters of order N = 4.

742

−45 −90 Bessel

−135 −180

Butterworth



Design of Digital Filters

and phase responses of a Bessel filter and Butterworth filter of order N = 4. We note that the Bessel filter has a larger transition bandwidth, but its phase is linear within the passband. However, we should emphasize that the linear-phase chacteristics of the analog filter are destroyed in the process of converting the filter into the digital domain by means of the transformations described previously.

3.5

Some Examples of Digital Filter Designs Based on the Bilinear Transformation

In this section we present several examples of digital filter designs obtained from analog filters by applying the bilinear transformation to convert H (s) to H (z). These filter designs are performed with the aid of one of several software packages now available for use on a personal computer. A lowpass filter is designed to meet specifications of a maximum ripple of 21 dB in the passband, 60-dB attenuation in the stopband, a passband edge frequency of ωp = 0.25π , and a stopband edge frequency of ωs = 0.30π . A Butterworth filter of order N = 37 is required to satisfy the specifications. Its frequency response characteristics are illustrated in Fig. 3.16. If a Chebyshev filter

Figure 3.16 Frequency response characteristics of a 37-order Butter-

worth filter.

743

Design of Digital Filters

is used, a filter of order N = 13 satisfies the specifications. The frequency response characteristics for a type I Chebyshev filter are shown in Fig. 3.17. The filter has a passband ripple of 0.31 dB. Finally, an elliptic filter of order N = 7 is designed which also satisfies the specifications. For illustrative purposes, we show in Table 6, the numerical values for the filter parameters, and the resulting frequency specifications are shown in Fig. 3.18. The following notation is used for the parameters in the function H (z): H (z) =

K % b(i, 0) + b(i, 1)z−1 + b(i, 2)z−2 i=1

1 + a(i, 1)z−1 + a(i, 2)z−2

(3.71)

Although we have described only lowpass analog filters in the preceding section, it is a simple matter to convert a lowpass analog filter into a bandpass, bandstop, or highpass analog filter by a frequency transformation, as is described in Section 4. The bilinear transformation is then applied to convert the analog filter into an equivalent digital filter. As in the case of the lowpass filters described above, the entire design can be carried out on a computer.

Figure 3.17 Frequency response characteristics of a 13-order type I

Chebyshev filter.

744

Design of Digital Filters

TABLE 6 Filter Coefficients for a 7-Order Elliptic Filter INFINITE IMPULSE RESPONSE (IIR) ELLIPTIC LOWPASS FILTER UNQUANTIZED COEFFICIENTS FILTER ORDER = 7 SAMPLING FREQUENCY = 2.000 KILOHERTZ I. A(I, 1) A(I, 2) B(I, 0) B(I, 1) B(I, 2) 1 -.790103 .000000 .104948 .104948 .000000 2 -1.517223 .714088 .102450 -.007817 .102232 3 -1.421773 .861895 .420100 -.399842 .419864 4 -1.387447 .962252 .714929 -.826743 .714841 *** CHARACTERISTICS OF DESIGNED FILTER *** BAND 1 BAND 2 LOWER BAND EDGE .00000 .30000 UPPER BAND EDGE .25000 1.00000 NOMINAL GAIN 1.00000 .00000 NOMINAL RIPPLE .05600 .00100 MAXIMUM RIPPLE .04910 .00071 RIPPLE IN DB .41634 -63.00399

Figure 3.18 Frequency response characteristics of a 7-order elliptic

filter.

745

Design of Digital Filters

4 Frequency Transformations The treatment in the preceding section is focused primarily on the design of lowpass IIR filters. If we wish to design a highpass or a bandpass or a bandstop filter, it is a simple matter to take a lowpass prototype filter (Butterworth, Chebyshev, elliptic, Bessel) and perform a frequency transformation. One possibility is to perform the frequency transformation in the analog domain and then to convert the analog filter into a corresponding digital filter by a mapping of the s -plane into the z-plane. An alternative approach is first to convert the analog lowpass filter into a lowpass digital filter and then to transform the lowpass digital filter into the desired digital filter by a digital transformation. In general, these two approaches yield different results, except for the bilinear transformation, in which case the resulting filter designs are identical. These two approaches are described below.

4.1

Frequency Transformations in the Analog Domain

First, we consider frequency transformations in the analog domain. Suppose that we have a lowpass filter with passband edge frequency p and we wish to convert it to another lowpass filter with passband edge frequency p . The transformation that accomplishes this is s −→

p s, p

(lowpass to lowpass)

(4.1)

Thus we obtain a lowpass filter with system function Hl (s) = Hp [( p / p )s], where Hp (s) is the system function of the prototype filter with passband edge frequency p . If we wish to convert a lowpass filter into a highpass filter with passband edge frequency p , the desired transformation is s −→

p p s

,

(lowpass to highpass)

(4.2)

The system function of the highpass filter is Hh (s) = Hp ( p p /s). The transformation for converting a lowpass analog filter with passband edge frequency c into a band filter, having a lower band edge frequency l and an upper band edge frequency u , can be accomplished by first converting the lowpass filter into another lowpass filter having a band edge frequency p = 1 and then performing the transformation s −→

s 2 + l u , s( u − l )

(lowpass to bandpass)

(4.3)

Equivalently, we can accomplish the same result in a single step by means of the transformation s −→ p

746

s 2 + l u , s( u − l )

(lowpass to bandpass)

(4.4)

Design of Digital Filters

TABLE 7 Frequency Transformations for Analog Filters (Pro-

totype Lowpass Filter Has Band Edge Frequency p ) Type of transformation

Transformation s −→

Lowpass

Band edge frequencies of new filter

p s p

p

p p

p

Highpass

s −→

Bandpass

s −→ p

s 2 + l u s( u − l )

l , u

Bandstop

s −→ p

s( u − c ) s 2 + u l

l , u

s

where l = lower band edge frequency u = upper band edge frequency Thus we obtain

 s 2 + l u Hb (s) = Hp p s( u − l )

Finally, if we wish to convert a lowpass analog filter with band-edge frequency p into a bandstop filter, the transformation is simply the inverse of (4.3) with the additional factor p serving to normalize for the band-edge frequency of the lowpass filter. Thus the transformation is s −→ p

s( u − l ) , s 2 + u l

which leads to

(lowpass to bandstop)

 Hbs (s) = Hp

s( u − l ) p 2 s + u l

(4.5)



The mappings in (4.1), (4.2), (4.3), and (4.5) are summarized in Table 7. The mappings in (4.4) and (4.5) are nonlinear and may appear to distort the frequency response characteristics of the lowpass filter. However, the effects of the nonlinearity on the frequency response are minor, primarily affecting the frequency scale but preserving the amplitude response characteristics of the filter. Thus an equiripple lowpass filter is transformed into an equiripple bandpass or bandstop or highpass filter.

747

Design of Digital Filters

EXAMPLE 4.1 Transform the single-pole lowpass Butterworth filter with system function H (s) =

p s + p

into a bandpass filter with upper and lower band edge frequencies u and l , respectively. Solution.

The desired transformation is given by (4.4). Thus we have H (s) =

=

1 s + l u +1 s( u − l ) 2

( u − l )s s 2 + ( u − l )s + l u

The resulting filter has a zero at s = 0 and poles at ( −( u − l ) ± 2u + 2l − 6 u l s= 2

4.2

Frequency Transformations in the Digital Domain

As in the analog domain, frequency transformations can be performed on a digital lowpass filter to convert it to either a bandpass, bandstop, or highpass filter. The transformation involves replacing the variable z−1 by a rational function g(z−1 ), which must satisfy the following properties: 1. The mapping z−1 −→ g(z−1 ) must map points inside the unit circle in the z-plane into itself. 2. The unit circle must also be mapped into itself. Condition (2) implies that for r = 1, e−j ω = g(e−j ω ) ≡ g(ω) = |g(ω)|ej arg[g(ω)] It is clear that we must have |g(ω)| = 1 for all ω. That is, the mapping must be all pass. Hence it is of the form g(z−1 ) = ±

n % z−1 − ak 1 − ak z−1

(4.6)

k=1

where |ak | < 1 to ensure that a stable filter is transformed into another stable filter (i.e., to satisfy condition 1). From the general form in (4.6), we obtain the desired set of digital transformations for converting a prototype digital lowpass filter into either a bandpass, a bandstop, a highpass, or another lowpass digital filter. These transformations are tabulated in Table 8.

748

Design of Digital Filters

TABLE 8 Frequency Transformation for Digital Filters (Prototype Lowpass Filter Has Band Edge Frequency ωp )

Type of transformation

Transformation

Lowpass

z−1 −→

Highpass

−1

Bandpass

Bandstop

z

Parameters ωp = band edge frequency new filter sin[(ωp − ωp )/2] a= sin[(ωp + ωp )/2]

z−1 − a 1 − az−1

z−1 + a −→ − 1 + az−1

z−1 −→ −

z−1 −→

z−2 − a1 z−1 + a2 a2 z−2 − a1 z−1 + 1

z−2 − a1 z−1 + a2 a2 z−1 − a1 z−1 + 1

ωp = band edge frequency new filter cos[(ωp + ωp )/2] a=− cos[(ωp − ωp )/2] = lower band edge frequency = upper band edge frequency = 2αK/(K + 1) = (K − 1)/(K + 1) cos[(ωu + ωl )/2] α = cos[(ωu − ωl )/2] ωp ωu − ωl K = cot tan 2 2

ωl ωu a1 a2

ωl = lower band edge frequency ωu = upper band edge frequency a1 = 2α/(K + 1) a2 = (1 − K)/(1 + K) cos[(ωu + ωl )/2] α = cos[(ωu − ωl )/2] ωp ωu − ωl K = tan tan 2 2

EXAMPLE 4.2 Convert the single-pole lowpass Butterworth filter with system function

H (z) =

0.245(1 + z−1 ) 1 − 0.509z−1

into a bandpass filter with upper and lower cutoff frequencies ωu and ωl , respectively. The lowpass filter has 3-dB bandwidth, ωp = 0.2π (see Example 3.5). Solution.

The desired transformation is z−1 −→ −

z−2 − a1 z−1 + a2 a2 z−2 − a1 z−1 + 1

749

Design of Digital Filters

where a1 and a2 are defined in Table 8. Substitution into H (z) yields   z−2 − a1 z−1 + a2 0.245 1 − a2 z−2 − a1 z−1 + 1 H (z) =  −2 z − a1 z−1 + a2 1 + 0.509 a2 z−2 − a1 z−1 + 1 =

0.245(1 − a2 )(1 − z−2 ) (1 + 0.509a2 ) − 1.509a1 z−1 + (a2 + 0.509)z−2

Note that the resulting filter has zeros at z = ±1 and a pair of poles that depend on the choice of ωu and ωl . For example, suppose that ωu = 3π/5 and ωl = 2π/5. Since ωp = 0.2π , we find that K = 1, a2 = 0, and a1 = 0. Then H (z) =

0.245(1 − z−2 ) 1 + 0.509z−2

This filter has poles at z = ±j 0.713 and hence resonates at ω = π/2.

Since a frequency transformation can be performed either in the analog domain or in the digital domain, the filter designer has a choice as to which approach to take. However, some caution must be exercised depending on the types of filters being designed. In particular, we know that the impulse invariance method and the mapping of derivatives are inappropriate to use in designing highpass and many bandpass filters, due to the aliasing problem. Consequently, one would not employ an analog frequency transformation followed by conversion of the result into the digital domain by use of these two mappings. Instead, it is much better to perform the mapping from an analog lowpass filter into a digital lowpass filter by either of these mappings, and then to perform the frequency transformation in the digital domain. Thus the problem of aliasing is avoided. In the case of the bilinear transformation, where aliasing is not a problem, it does not matter whether the frequency transformation is performed in the analog domain or in the digital domain. In fact, in this case only, the two approaches result in identical digital filters.

5 Summary and References We have described in some detail the most important techniques for designing FIR and IIR digital filters based either on frequency-domain specifications expressed in terms of a desired frequency response Hd (ω) or on the desired impulse response hd (n). As a general rule, FIR filters are used in applications where there is a need for a linear-phase filter. This requirement occurs in many applications, especially in telecommunications, where there is a requirement to separate (demultiplex) signals such as data that have been frequency-division multiplexed, without distorting these

750

Design of Digital Filters

signals in the process of demultiplexing. Of the several methods described for designing FIR filters, the frequency-sampling design method and the optimum Chebyshev approximation method yield the best designs. IIR filters are generally used in applications where some phase distortion is tolerable. Of the class of IIR filters, elliptic filters are the most efficient to implement in the sense that for a given set of specifications, an elliptic filter has a lower order or fewer coefficients than any other IIR filter type. When compared with FIR filters, elliptic filters are also considerably more efficient. In view of this, one might consider the use of an elliptic filter to obtain the desired frequency selectivity, followed then by an all-pass phase equalizer that compensates for the phase distortion in the elliptic filter. However, attempts to accomplish this have resulted in filters with a number of coefficients in the cascade combination that equaled or exceeded the number of coefficients in an equivalent linear-phase FIR filter. Consequently, no reduction in complexity is achievable in using phase-equalized elliptic filters. Such a rich literature now exists on the design of digital filters that it is not possible to cite all the important references. We shall cite only a few. Some of the early work on digital filter design was done by Kaiser (1963, 1966), Steiglitz (1965), Golden and Kaiser (1964), Rader and Gold (1967a), Shanks (1967), Helms (1968), Gibbs (1969, 1970), and Gold and Rader (1969). The design of analog filters is treated in the classic books by Storer (1957), Guillemin (1957), Weinberg (1962), and Daniels (1974). The frequency-sampling method for filter design was first proposed by Gold and Jordan (1968, 1969), and optimized by Rabiner et al. (1970). Additional results were published by Herrmann (1970), Herrmann and Schuessler (1970a), and Hofstetter et al. (1971). The Chebyshev (minimax) approximation method for designing linearphase FIR filters was proposed by Parks and McClellan (1972a,b) and discussed further by Rabiner et al. (1975). The design of elliptic digital filters is treated in the book by Gold and Rader (1969) and in the paper by Gray and Markel (1976). The latter includes a computer program for designing digital elliptic filters. The use of frequency transformations in the digital domain was proposed by Constantinides (1967, 1968, 1970). These transformations are appropriate only for IIR filters. The reader should note that when these transformations are applied to a lowpass FIR filter, the resulting filter is IIR. Direct design techniques for digital filters have been considered in a number of papers, including Shanks (1967), Burrus and Parks (1970), Steiglitz (1970), Deczky (1972), Brophy and Salazar (1973), and Bandler and Bardakjian (1973).

Problems 1

Design an FIR linear-phase, digital filter approximating the ideal frequency response  π  1, for |ω| ≤ 6 Hd (ω) = π  0, for < |ω| ≤ π 6 (a) Determine the coefficients of a 25-tap filter based on the window method with a rectangular window.

751

Design of Digital Filters

(b) Determine and plot the magnitude and phase response of the filter. (c) Repeat parts (a) and (b) using the Hamming window. (d) Repeat parts (a) and (b) using a Bartlett window. 2 Repeat Problem 1 for a bandstop filter having the ideal response   1,     Hd (ω) = 0,      1, 3 4 5

for |ω| ≤

π 6

π π < |ω| < 6 3 π for ≤ |ω| ≤ π 3

for

Redesign the filter of Problem 1 using the Hanning and Blackman windows. Redesign the filter of Problem 2 using the Hanning and Blackman windows. Determine the unit sample response {h(n)} of a linear-phase FIR filter of length M = 4 for which the frequency response at ω = 0 and ω = π/2 is specified as Hr (0) = 1,

Hr

π  2

=

1 2

6

Determine the coefficients {h(n)} of a linear-phase FIR filter of length M = 15 which has a symmetric unit sample response and a frequency response that satisfies the condition   2π k 1, k = 0, 1, 2, 3 Hr = 0, k = 4, 5, 6, 7 15

7

Repeat the filter design problem in Problem 6 with the frequency response spec ifications   1, k = 0, 1, 2, 3 2π k Hr = 0.4, k=4 15 0, k = 5, 6, 7

8

The ideal analog differentiator is described by ya (t) =

dxa (t) dt

where xa (t) is the input and ya (t) the output signal. (a) Determine its frequency response by exciting the system with the input xa (t) = ej 2π F t . (b) Sketch the magnitude and phase response of an ideal analog differentiator bandlimited to B hertz. (c) The ideal digital differentiator is defined as H (ω) = j ω,

|ω| ≤ π

Justify this definition by comparing the frequency response |H (ω)|, ⭿H (ω) with that in part (b).

752

Design of Digital Filters

(d) By computing the frequency response H (ω), show that the discrete-time system y(n) = x(n) − x(n − 1) is a good approximation of a differentiator at low frequencies. (e) Compute the response of the system to the input x(n) = A cos(ω0 n + θ) 9

Use the window method with a Hamming window to design a 21-tap differentiator as shown in Fig. P9. Compute and plot the magnitude and phase response of the resulting filter. |Hd (ω)|

π

Figure P9

10

−π

0

π

ω

Use the bilinear transformation to convert the analog filter with system function H (s) =

s + 0.1 (s + 0.1)2 + 9

into a digital IIR filter. Select T = 0.1 and compare the location of the zeros in H (z) with the locations of the zeros obtained by applying the impulse invariance method in the conversion of H (s). 11 Convert the analog bandpass filter designed in Example 4.1 into a digital filter by means of the bilinear transformation. Thereby derive the digital filter characteristic obtained in Example 4.2 by the alternative approach and verify that the bilinear transformation applied to the analog filter results in the same digital bandpass filter. 12 An ideal analog integrator is described by the system function Ha (s) = 1/s . A digital integrator with system function H (z) can be obtained by use of the bilinear transformation. That is, H (z) =

T 1 + z−1 ≡ Ha (s)|s=(2/T )(1−z−1 )/(1+z−1 ) 2 1 − z−1

(a) Write the difference equation for the digital integrator relating the input x(n) to the output y(n). (b) Roughly sketch the magnitude |Ha (j )| and phase ( ) of the analog integrator.

753

Design of Digital Filters

(c) It is easily verified that the frequency response of the digital integrator is H (ω) = −j

T cos(ω/2) T ω = −j cot 2 sin(ω/2) 2 2

Roughly sketch |H (ω)| and θ(ω). (d) Compare the magnitude and phase characteristics obtained in parts (b) and (c). How well does the digital integrator match the magnitude and phase characteristics of the analog integrator? (e) The digital integrator has a pole at z = 1. If you implement this filter on a digital computer, what restrictions might you place on the input signal sequence x(n) to avoid computational difficulties? 13

A z-plane pole–zero plot for a certain digital filter is shown in Fig. P13. The filter has unity gain at dc.

|z| =

1 2

3 zeros @ z = −1

|z| = 1

60

−60

Figure P13

(a) Determine the system function in the form !

(1 + a1 z−1 )(1 + b1 z−1 + b2 z−2 ) H (z) = A (1 + c1 z−1 )(1 + d1 z−1 + d2 z−2 )

"

giving numerical values for the parameters A, a1 , b1 , b2 , c1 , d1 , and d2 . (b) Draw block diagrams showing numerical values for path gains in the following forms: (a) Direct form II (canonic form) (b) Cascade form (make each section canonic, with real coefficients)

754

Design of Digital Filters

14

Consider the pole–zero plot shown in Fig. P14.

|z| =

4 3 |z| = 1 |z| =

3 4

60 −60

Figure P14

(a) Does it represent an FIR filter? (b) Is it a linear-phase system? (c) Give a direct-form realization that exploits all symmetries to minimize the number of multiplications. Show all path gains. 15 A digital low-pass filter is required to meet the following specifications: Passband ripple: ≤ 1 dB Passband edge: 4 kHz Stopband attenuation: ≥ 40 dB Stopband edge: 6 kHz Sample rate: 24 kHz

16

The filter is to be designed by performing a bilinear transformation on an analog system function. Determine what order Butterworth, Chebyshev, and elliptic analog designs must be used to meet the specifications in the digital implementation. An IIR digital lowpass filter is required to meet the following specifications: Passband ripple (or peak-to-peak ripple): ≤ 0.5 dB Passband edge: 1.2 kHz Stopband attenuation: ≥ 40 dB Stopband edge: 2.0 kHz Sample rate: 8.0 kHz Use the design formulas in the book to determine the required filter order for (a) A digital Butterworth filter (b) A digital Chebyshev filter (c) A digital elliptic filter

755

Design of Digital Filters

17

Determine the system function H (z) of the lowest-order Chebyshev digital filter that meets the following specifications: (a) 1-dB ripple in the passband 0 ≤ |ω| ≤ 0.3π . (b) At least 60 dB attentuation in the stopband 0.35π ≤ |ω| ≤ π . Use the bilinear transformation.

18 Determine the system function H (z) of the lowest-order Chebyshev digital filter that meets the following specifications: (a)

1 2 -dB

ripple in the passband 0 ≤ |ω| ≤ 0.24π .

(b) At least 50-dB attenuation in the stopband 0.35π ≤ |ω| ≤ π . Use the bilinear transformation. 19 An analog signal x(t) consists of the sum of two components x1 (t) and x2 (t). The spectral characteristics of x(t) are shown in the sketch in Fig. P19. The signal x(t) is bandlimited to 40 kHz and it is sampled at a rate of 100 kHz to yield the sequence x(n). |X(F)| Spectrum of x1(t) Spectrum of x2(t)

Figure P19

0

20

40

F

Frequency in kilohertz

It is desired to suppress the signal x2 (t) by passing the sequence x(n) through a digital lowpass filter. The allowable amplitude distortion on |X1 (f )| is ±2%(δ1 = 0.02) over the range 0 ≤ |F | ≤ 15 kHz. Above 20 kHz, the filter must have an attenuation of at least 40 dB (δ2 = 0.01). (a) Use the Remez exchange algorithm to design the minimum-order linear-phase FIR filter that meets the specifications above. From the plot of the magnitude characteristic of the filter frequency response, give the actual specifications achieved by the filter. (b) Compare the order M obtained in part (a) with the approximate formulas given in equations (2.94) and (2.95). (c) For the order M obtained in part (a), design an FIR digital lowpass filter using the window technique and the Hamming window. Compare the frequency response characteristics of this design with those obtained in part (a). (d) Design the minimum-order elliptic filter that meets the given amplitude specifications. Compare the frequency response of the elliptic filter with that of the FIR filter in part (a).

756

Design of Digital Filters

(e) Compare the complexity of implementing the FIR filter in part (a) versus the elliptic filter obtained in part (d). Assume that the FIR filter is implemented in the direct form and the elliptic filter is implemented as a cascade of two-pole filters. Use storage requirements and the number of multiplications per output point in the comparison of complexity. 20

The impulse response of an analog filter is shown in Fig. P20.

ha(t) 5

Figure P20 0

5

10

t

(a) Let h(n) = ha (nT ), where T = 1, be the impulse response of a discrete-time filter. Determine the system function H (z) and the frequency response H (ω) for this FIR filter. (b) Sketch (roughly) |H (ω) and compare this frequency response characteristic with |Ha (j )|. 21

In this problem you will be comparing some of the characteristics of analog and digital implementations of the single-pole low-pass analog system Ha (s) =

α ⇔ ha (t) = e−αt s+α

(a) What is the gain at dc? At what radian frequency is the analog frequency response 3 dB down from its dc value? At what frequency is the analog frequency response zero? At what time has the analog impulse response decayed to 1/e of its initial value? (b) Give the digital system function H (z) for the impulse-invariant design for this filter. What is the gain at dc? Give an expression for the 3-dB radian frequency. At what (real-valued) frequency is the response zero? How many samples are there in the unit sample time-domain response before it has decayed to 1/e of its initial value? (c) “Prewarp” the parameter α and perform the bilinear transformation to obtain the digital system function H (z) from the analog design. What is the gain at dc? At what (real-valued) frequency is the response zero? Give an expression for the 3-dB radian frequency. How many samples are there in the unit sample time-domain response before it has decayed to 1/e of its initial value?

757

Design of Digital Filters

22

We wish to design a FIR bandpass filter having a duration M = 201. Hd (ω) represents the ideal characteristic of the noncausal bandpass filter as shown in Fig. P22. Hd (ω)

−π

−0.5π

−0.4π

−0.4π

0

π

0.5π

ω

Figure P22

(a) Determine the unit sample (impulse) response hd (n) corresponding to Hd (ω). (b) Explain how you would use the Hamming window 2π n , w(n) = 0.54 + 0.46 cos N −1 



M −1 M −1 ≤n≤ 2 2

to design a FIR bandpass filter having an impulse response h(n) for 0 ≤ n ≤ 200. (c) Suppose that you were to design the FIR filter with M = 201 by using the frequency-sampling technique in which the DFT coefficients H (k) are specified instead of h(n). Give the values of H (k) for 0 ≤ k ≤ 200 corresponding to Hd (ej ω ) and indicate how the frequency response of the actual filter will differ from the ideal. Would the actual filter represent a good design? Explain your answer. 23

We wish to design a digital bandpass filter from a second-order analog lowpass Butterworth filter prototype using the bilinear transformation. The specifications on the digital filter are shown in Fig. P23 (a). The cutoff frequencies (measured at the half-power points) for the digital filter should lie at ω = 5π/12 and ω = 7π/12. The analog prototype is given by H (s) =

1 √ s 2 + 2s + 1

with the half-power point at = 1. (a) Determine the system function for the digital bandpass filter. (b) Using the same specs on the digital filter as in part (a), determine which of the analog bandpass prototype filters shown in Fig. P23 (b) could be transformed directly using the bilinear transformation to give the proper digital filter. Only the plot of the magnitude squared of the frequency is given.

758

Design of Digital Filters

|H(ω)| 2

π 6

1 1 power points 2 π 2

0

π

ω

(a)

|H(Ω)| 2

1 I.

1 2 200



285

|H(Ω)| 2

1 II.

1 2 111



200

|H(Ω)| 2

1 III.

1 2 300



547

|H(Ω)| 2

1 IV.

1 2 600

1019



(b)

Figure P23

759

Design of Digital Filters

24

Figure P24 shows a digital filter designed using the frequency-sampling method. +

1 6 z −1 −1

+

x(n)

+ −

+

+

1 12

+

ν(n)

z −1

z −12

1 −2

1

+

z −1 −1

+

+

1 12

z −1 1 2

−1

+

z −1 −1

Figure P24

(a) Sketch a z-plane pole–zero plot for this filter. (b) Is the filter lowpass, highpass, or bandpass? (c) Determine the magnitude response |H (ω)| at the frequencies ωk = π k/6 for k = 0, 1, 2, 3, 4, 5, 6. (d) Use the results of part (c) to sketch the magnitude response for 0 ≤ ω ≤ π and confirm your answer to part (b). 25 An analog signal of the form xa (t) = a(t) cos 2000π t is bandlimited to the range 900 ≤ F ≤ 1100 Hz. It is used as an input to the system shown in Fig. P25. A/D converter

xa(t)

Rx = 1 = 2500 Tx

Figure P25

760

x(n) X

ω(n)

cos (0.8 πn)

H(ω)

ν(n)

D/A converter

â(t)

Design of Digital Filters

(a) Determine and sketch the spectra for the signals x(n) and w(n). (b) Use a Hamming window of length M = 31 to design a lowpass linear phase FIR filter H (ω) that passes {a(n)}. (c) Determine the sampling rate of the A/D converter that would allow us to eliminate the frequency conversion in Fig. P25. 26 System identification Consider an unknown LTI system and an FIR system model as shown in Fig. P26. Both systems are excited by the same input sequence {x(n)}. The problem is to determine the coefficients {h(n), 0 ≤ n ≤ M − 1} of the FIR model of the system to minimize the average squared error between the outputs of the two systems. Unknown LTI system

y(n)

+

x(n)



+

ˆy(n)

FIR model

Minimize the sum of squared errors

Figure P26

(a) Determine the equation for the FIR filter coefficients {h(n), 0 ≤ n ≤ M − 1} that minimize the least-squares error E=

N 

2 [y(n) − y(n)] ˆ

n=0

where y(n) ˆ =

M−1 

h(k)x(n − k),

n = K, K + 1, . . . , N

k=0

and N  M . (b) Repeat part (a) if the output of the unknown system is corrupted by an additive white noise {w(n)} sequence with variance σw2 .

761

Design of Digital Filters

27

A linear time-invariant system has an input sequence x(n) and an output sequence y(n). The user has access only to the system output y(n). In addition, the following information is available: The input signal is periodic with a given fundamental period N and has a flat spectral envelope, that is, x(n) =

N−1 

ckx ej (2π/N)kn ,

all n

k=0

where ckx = 1 for all k. The system H (z) is all pole, that is, 1

H (z) = 1+

P 

ak z−k

k=1

but the order p and the coefficients (ak , 1 ≤ k ≤ p) are unknown. Is it possible to determine the order p and the numerical values of the coefficients {ak , 1 ≤ k ≤ p} by taking measurements on the output y(n)? If yes, explain how. Is this possible for every value of p? 28 FIR system modeling Consider an “unknown” FIR system with impulse response h(n), 0 ≤ n ≤ 11, given by h(0) = h(11) = 0.309828 × 10−1 h(1) = h(10) = 0.416901 × 10−1 h(2) = h(9) = −0.577081 × 10−1 h(3) = h(8) = −0.852502 × 10−1 h(4) = h(7) = 0.147157 × 100 h(5) = h(6) = 0.449188 × 100 A potential user has access to the input and output of the system but does not have any information about its impulse response other than that it is FIR. In an effort to determine the impulse response of the system, the user excites it with a zero-mean, random sequence x(n) uniformly distributed in the range [−0.5, 0.5], and records the signal x(n) and the corresponding output y(n) for 0 ≤ n ≤ 199. (a) By using the available information that the unknown system is FIR, the user employs the method of least squares to obtain an FIR model h(n), 0 ≤ n ≤ M − 1. Set up the system of linear equations, specifying the parameters h(0), h(1), . . . , h(M − 1). Specify formulas we should use to determine the necessary autocorrelation and crosscorrelation values.

762

Design of Digital Filters

(b) Since the order of the system is unknown, the user decides to try models of different orders and check the corresponding total squared error. Clearly, this error will be zero (or very close to it if the order of the model becomes equal to the order of the system). Compute the FIR models hM (n), 0 ≤ n ≤ M − 1 for M = 8, 9, 10, 11, 12, 13, 14 as well as the corresponding total squared errors EM , M = 8, 9, . . . , 14. What do you observe? (c) Determine and plot the frequency response of the system and the models for M = 11, 12, 13. Comment on the results. (d) Suppose now that the output of the system is corrupted by additive noise, so instead of the signal y(n), 0 ≤ n ≤ 199, we have available the signal v(n) = y(n) + 0.01w(n) where w(n) is a Gaussian random sequence with zero mean and variance σ 2 = 1. Repeat part (b) of Problem 27 by using v(n) instead of y(n) and comment on the results. The quality of the model can be also determined by the quantity ∞ 

Q=

2 ˆ [h(n) − h(n)]

n=0 ∞ 

h2 (n)

n=0

29 Filter design by Padé approximation Let the desired impulse response hd (n), n ≥ 0, of an IIR filter be specified. The filter to be designed to approximate {hd (n)} has the system function M ∞ −k  k=0 bk z = h(k)z−k H (z) = N 1 + k=1 ak z−k k=0 H (z) has L = M + N + 1 parameters, namely, the coefficients {ak } and {bk } to be determined. Suppose the input to the filter is x(n) = δ(n). Then, the response of the filter is y(n) = h(n), and, hence, h(n) = −a1 h(n − 1) − a2 h(n − 2) − . . . − aN h(n − N ) + b0 δ(n) + b1 δ(n − 1) + . . . + bM δ(n − M)

(1)

(a) Show that equation (1) reduces to h(n) = −a1 h(n − 1) − a2 h(n − 2) − . . . − aN h(n − N ) + bn ,

0≤n≤M

(2)

n>M

(3)

(b) Show that for n > M , equation (1) reduces to h(n) = −a1 h(n − 1) − a2 h(n − 2) − . . . − aN h(n − N),

(c) Explain how equations (2) and (3) can be used to determine {ak } and {bk } by letting h(n) = hd (n) for 0 ≤ n ≤ N + M . (This filter design method in which h(n) exactly matches the desired response hd (n) for 0 ≤ n ≤ M + N is called the Padé approximation method.)

763

Design of Digital Filters

30

Suppose the desired unit sample response is  n 1 hd (n) = 2 u(n) 2 (a) Use the Padé approximation described in Problem 29 to determine h(n). (b) Compare the frequency response of H (ω) with that of the desired filter response Hd (ω).

31 Shanks method for least-squares filter design Suppose that we are given the desired response hd (n), n ≥ 0, and we wish to determine the coefficients {ak } and {bk } of an IIR filter with system function M H (z) =

k=0

N

1+

bk z−k

k=1 ak z

−k

such that the sum of squared errors between hd (n) and h(n) is minimized. (a) If the input to H (z) is x(n) = δ(n), what is the difference equation satisfied by the filter H (z)? (b) Show that for n > M , an estimate of hd (n) is hˆ d (n) = −

N 

ak hd (n − k)

k=1

and determine the equations for the coefficients {ak } by minimizing the sum of squared errors ∞  [hd (n) − hˆ d (n)]2 E1 = n=M+1

Thus, the filter coefficients {ak } that specify the denominator of H (z) are determined. (c) To determine the parameters {bk } consider the system shown in Fig. P31, where H1 (z) =

H2 (z) =

1+ N 

1 N

ˆkz k=1 a

−k

bk z−k

k=1

and {aˆ k } are the coefficients determined in part (b). If the response of H1 (z) to the input δ(n) is denoted as v(n) = −

N  k=1

764

aˆ k v(n − k) + δ(n)

Design of Digital Filters

δ(n)

ν(n)

all-pole filter

All-zero filter

H1(z)

hd(n)

H2(z)

Figure P31

and the output of H2 (z) is denoted as hˆ d (n), determine the equation for the parameters {bk } that minimize the sum of squared errors E2 =

∞ 

[hd (n) − hˆ d (n)]2

n=0

(The preceding least-squares method for filter design is due to Shanks (1967).) 32 Use the Shanks method for filter design as described in Problem 31 to determine the parameters {ak } and {bk } of M H (z) =

k=0

1+

N

bk z−k

k=1 ak z

−k

when the desired response is the impulse response of the three-pole and three-zero type II lowpass Chebyshev digital filter having a system function Hd (z) =

0.3060(1 + z−1 )(0.2652 − 0.09z−1 + 0.2652z−2 ) (1 − 0.3880z−1 )(1 − 1.1318z−1 + 0.5387z−2 )

(a) Plot hd (n) and then observe that hd (n) ≈ 0 for n > 50. (b) Determine the pole and zero positions for the filter H (z) obtained by the Shanks method for (N, M) = (3, 2), (N, M) = (3, 3) and (N, M) = (4, 3), and compare these results with the poles and zeros of Hd (z). Comment on the similarities and differences.

Answers to Selected Problems sin π (n−12)

1

6 hd (n) = π(n−12) ; h(n) = hd (n)w(n) where w(n) is a rectangular window of length N = 25.

2

3 6 + π(n−12) ; h(n) = hd (n)w(n) hd (n) = δ(n) − π(n−12) where w(n) is a rectangular window of length 25.

   Hr (ω) = 2 1n=0 h(n) cos ω 23 − n π  From Hr (0) = 1 and Hr 2 = 1/2, we obtain h(0) = 0.073, h(1) = 0.427, h(2) = h(1) and h(3) = h(0) , h(n) = {0.3133, −0.0181, −0.0914, 0.0122, 0.0400, −0.0019, −0.0141, 0.52, 0.52, −0.0141 −0.0019, 0.0400, 0.0122, −0.0914, −0.0181, 0.3133}

5

7

9

sin π (n−2)

hd (n) =

cos π(n − 10) , (n − 10)

sin π (n−12)

0 ≤ n ≤ 20, n = 10

= 0, n = 10 Then h(n) = hd (n)w(n), where w(n) is the Hamming window of length 21. 12

(a) Let T = 2. Then H (z) =

13

H (z) = A 

1+z−1 1−z−1

⇒ y(n) = y(n − 1) + x(n) + x(n − 1)



  1+z−1 1+2z−1 +z−2   1 1 1 −1 −1 −2 1− 2 z 1− 2 z + 4 z

H (z)|z=1 = 1; A =

3 64

, b1 = 2, b2 = 1, a1 = 1, c1 = − 21 , d1 = − 21 , d2 =

1 4

765

Design of Digital Filters

15 From the design specifications we obtain ε = 0.509, δ = 99.995, fp = 16 , fs = Butterworth filter: Nmin ≥ Chebyshev filter: Nmin ≥ Elliptic filter: Nmin ≥

1 4

log η = 9.613 ⇒ N = 10 log k cos h−1 η = 5.212 ⇒ N = 6 cos h−1 k k(sin α) k(cos β) , = 3.78 ⇒ N = 4 k(cos α) k(sin β)

19 (a) MATLAB is used to design the FIR filter using the Remez algorithm. We find that a filter of length M = 37 meets the specifications. We note that in MATLAB, the frequency scale is normalized to 21 of the sampling frequency. (b) δ1 = 0.02, δ2 = 0.01, f =

= 0.05 √  −20 log10 δ1 δ2 −13 ˆ With equation (2.94) we obtain M = + 1 ≈ 34 14.6f With equation (2.95) we obtain D ∞ (δ1 δ2 ) = 1.7371; f (δ1 δ2 ) = 11.166 2 ˆ = D∞ (δ1 δ2 )−f (δ1 δ2 )(f ) + 1 ≈ 36 and M 20 100



15 100

f

Note (2.95) is a better approximation of M . 21 (a) dc gain: Ha (0) = 1; 3 dB frequency: c = α For all , only H (j ∞) = 0; ha (τ ) = 1e ha (0) = 1e ; r = α1   24 H (z) = 16 (1 − z−6 )(1 − z−1 ) 2 + z−1 + 23 z−α + 21 z−3 + z−4 π π π This filter is FIR with zeros at z = 1, e±j 6 , e±j 2 , e±j 5 6 , −0.55528±j 0.6823 and 0.3028±j 0.7462 900 = 0.36; fH = 1100 = 0.44 25 (a) fL = 2500 2500 2 sin 0.08π(n−15) hd (n) = ; h(n) = hd (n)wH (n) (n−15)

wH (n) = 0.54 − 0.46 cos

766

2π(n−15) 30

Multirate Digital Signal Processing

From Chapter 11 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

767

Multirate Digital Signal Processing

In many practical applications of digital signal processing, one is faced with the problem of changing the sampling rate of a signal, either increasing it or decreasing it by some amount. For example, in telecommunication systems that transmit and receive different types of signals (e.g., teletype, facsimile, speech, video, etc.), there is a requirement to process the various signals at different rates commensurate with the corresponding bandwidths of the signals. The process of converting a signal from a given rate to a different rate is called sampling rate conversion. In turn, systems that employ multiple sampling rates in the processing of digital signals are called multirate digital signal processing systems. Sampling rate conversion of a digital signal can be accomplished in one of two general methods. One method is to pass the digital signal through a D/A converter, filter it if necessary, and then to resample the resulting analog signal at the desired rate (i.e., to pass the analog signal through an A/D converter). The second method is to perform the sampling rate conversion entirely in the digital domain. One apparent advantage of the first method is that the new sampling rate can be arbitrarily selected and need not have any special relationship to the old sampling rate. A major disadvantage, however, is the signal distortion, introduced by the D/A converter in the signal reconstruction, and by the quantization effects in the A/D conversion. Sampling rate conversion performed in the digital domain avoids this major disadvantage. In this chapter we describe sampling rate conversion and multirate signal processing in the digital domain. First we describe sampling rate conversion by a rational factor and present several methods for implementing the rate converter, including single-stage and multistage implementations. Then, we describe a method for sampling rate conversion by an arbitrary factor and discuss its implementation. We

768

Multirate Digital Signal Processing

present several applications of sampling rate conversion in multirate signal processing systems, which include the implementation of narrowband filters, digital filter banks, subband coding, transmultiplexers, and quadrature mirror filters.

1

Introduction The process of sampling rate conversion can be developed and understood using the idea of “resampling after reconstruction.” In this theoretical approach, a discretetime signal is ideally reconstructed and the resulting continuous-time signal is resampled at a different sampling rate. This idea leads to a mathematical formulation that enables the realization of the entire process by means of digital signal processing. Let x(t) be a continuous-time signal that is sampled at a rate Fx = 1/Tx to generate a discrete-time signal x(nTx ). From the samples x(nTx ) we can generate a continuous-time signal using the interpolation formula y(t) =

∞ 

x(nTx )g(t − nTx )

(1.1)

n=−∞

If the bandwidth of x(t) is less than Fx /2 and the interpolation function is given by  sin(π t/Tx ) F Tx , |F | ≤ Fx /2 ←→ G(F ) = (1.2) g(t) = 0, otherwise π t/Tx then y(t) = x(t); otherwise y(t) = x(t). In practice, perfect recovery of x(t) is not possible because the infinite summation in (1.1) should be replaced by a finite summation. To perform sampling rate conversion we simply evaluate (1.1) at time instants t = mTy , where Fy = 1/Ty is the desired sampling frequency. Therefore, the general formula for sampling rate conversion becomes y(mTy ) =

∞ 

x(nTx )g(mTy − nTx )

(1.3)

n=−∞

which expresses directly the samples of the desired sequence in terms of the samples of the original sequence and sampled values of the reconstruction function at positions (mTy − nTx ). The computation of y(nTy ) requires (a) the input sequence x(nTx ), (b) the reconstruction function g(t), and (c) the time instants nTx and mTy of the input and output samples. The values y(mTy ) calculated by this equation are accurate only if Fy > Fx . If Fy < Fx , we should filter out the frequency components of x(t) above Fy /2 before resampling in order to prevent aliasing. Therefore, the sampling rate conversion formula (1.3) yields y(mTy ) = x(mTy ) if we use (1.2) and X(F ) = 0 for |F | ≥ min{Fx /2, Fy /2}. If Ty = Tx , equation (1.3) becomes a convolution summation, which corresponds to an LTI system. To understand the meaning of (1.3) for T y = Tx , we rearrange the argument of g(t) as follows:    ∞  mTy (1.4) x(nTx )g Tx −n y(mTy ) = Tx n=−∞

769

Multirate Digital Signal Processing

The term mTy /Tx can be decomposed into an integer part km and a fractional part m , 0 ≤ m < 1, as mTy = km + m (1.5) Tx where

 km =

and

mTy Tx

 (1.6)

  mTy mTy − m = Tx Tx

(1.7)

The symbol a denotes the largest integer contained in a . The quantity m specifies the position of the current sample within the sample period Tx . Substituting (1.5) into (1.4) we obtain ∞ 

y(mTy ) =

x(nTx )g((km + m − n)Tx )

(1.8)

n=−∞

If we change the index of summation in (1.8) from n to k = k

m

− n, we have

y(mTy ) = y((km + m )Tx ) =

∞ 

(1.9)

g(kTx + m Tx )x((km − k)Tx )

k=−∞

Equation (1.9) provides the fundamental equation for the discrete-time implementation of sampling rate conversion. This process is illustrated in Figure 1.1. We note that (a) given Tx and Ty the input and output sampling times are fixed, (b) the function g(t) is shifted for each m such that the value g(m Tx ) is positioned at t = mTy , and (c) the required values of g(t) are determined at the input sampling times. For each value of m, the fractional interval m determines the impulse response coefficients whereas the index km specifies the corresponding input samples  mT y  km =    Tx 

g(∆ mT x) Input sampling times

(k m −1) T x

(k m+1) T x (k m+2) T x

km Tx

x(t)

g(t) t Output sampling times

(m−1) T y

m Ty ∆mTx

(m+ 1) T y Fractional interval

Figure 1.1 Illustration of timing relations for sampling rate conversion.

770

Multirate Digital Signal Processing

Figure 1.2

Discrete-time linear time-varying system for sampling rate conversion.

x(nTx)

g(nTx + ∆mTx)

x(n)

g m(n)

y(mTy) y(m)

needed to compute the sample y(mTy ). Since for any given value of m the index km is an integer number, y(mTy ) is the convolution between the input sequence x(nTx ) and an impulse response g ((n + m ) Tx ). The difference between (1.8) and (1.9) is that the first shifts a “changing” reconstruction function whereas the second shifts the “fixed” input sequence. The sampling rate conversion process defined by (1.9) is a discrete-time linear and time-varying system, because it requires a different impulse response gm (nTx ) = g((n + m )Tx )

(1.10)

for each output sample y(mTy ). Therefore, a new set of coefficients should be computed or retrieved from storage for the computation of every output sample (see Figure 1.2). This procedure may be inefficient when the function g(t) is complicated and the number of required values is large. This dependence on m prohibits the use of recursive structures because the required past output values must be computed using an impulse response for the present value of m . A significant simplification results when the ratio Ty /Tx is constrained to be a rational number, that is, Ty Fx D = = (1.11) Fy I Tx where D and I are relatively prime integers. To this end, we express the offset m as       mD mD 1 mD 1 m = − = mD − I = (mD)I (1.12) I I I I I where (k)I denotes the value of k modulo I . From (1.12) it is clear that m can take on only I unique values 0, 1/I, . . . , (I − 1)/I , so that there are only I distinct impulse responses possible. Since gm (nTx ) can take on I distinct sets of values, it is periodic in m; that is, gm (nTx ) = gm+rI (nTx ),

r = 0, ±1, ±2, . . .

(1.13)

Thus the system gm (nTx ) is a linear and periodically time-varying discrete-time system. Such systems have been widely studied for a wide range of applications (Meyers and Burrus, 1975). This is a great simplification compared to the continuously timevarying discrete-time system in (1.10). To illustrate these concepts we consider two important special cases. We start with the process of reducing the sampling rate by an integer factor D, which is known as decimation or downsampling. If we set Ty = DTx in (1.3), we have y(mTy ) = y(mDTx ) =

∞ 

x(kTx )g((mD − k)Tx )

(1.14)

k=−∞

771

Multirate Digital Signal Processing

We note that the input signal and the impulse response are sampled with a period Tx . However, the impulse response is shifted at increments of Ty = DTx because we need to compute only one out of every D samples. Since I = 1, we have m = 0 and therefore there is only one impulse response g(nTx ), for all m. This process is illustrated in Figure 1.3 for D = 2. We consider now the process of increasing the sampling rate by an integer factor I , which is called upsampling or interpolation. If we set Ty = Tx /I in (1.3), we have ∞  x(kTx )g(m(Tx /I ) − kTx ) (1.15) y(mTy ) = k=−∞

We note that both x(t) and g(t) are sampled with a period Tx ; however, the impulse response is shifted at increments of Ty = Tx /I for the computation of each output sample. This is required to “fill-in” an additional number of (I − 1) samples within each period Tx . This is illustrated in Figure 1.4(a)–(b) for I = 2. Each “fractional shifting” requires that we resample g(t), resulting in a new impulse response gm (nTx ) = g(nTx +mTx /I ), for m = 0, 1, . . . , I −1 in agreement with (1.14). Careful inspection of Figure 1.4(a)–(b) shows that if we determine an impulse response sequence g(nTy ) and create a new sequence v(nTy ) by inserting (I − 1) zero samples between successive samples of x(nTx ), we can compute y(mTy ) as the convolution of the sequences g(nTy ) and x(nTy ). This idea is illustrated in Figure 1.4(c) for I = 2. In the next few sections we discuss in more detail the properties, design, and structures for the implementation of sampling rate conversion entirely in the discretetime domain. For convenience, we usually drop the sampling periods Tx and Ty from the argument of discrete-time signals. However, occasionally, it will be beneficial to the reader to reintroduce and think in terms of continuous-time quantities and units. t = mTy

t = (m + 1)Ty

x(t)

g(t)

(n − 1)Tx

nTx

(n + 1)Tx (n + 2)Tx

t y(t)

(m − 1)Ty

mTy

(m + 1)Ty

t

Figure 1.3 Illustration of timing relations for sampling rate decrease by an integer factor D = 2. A single impulse response, sampled with period Tx , is shifted at steps equal to Ty = DTx to generate the output samples.

772

Multirate Digital Signal Processing

t = m Ty

t = (m + 1) Ty

g0(nTx)

(a)

g1(nTx)

x(t) x(nTx)

g(t) (n−1)Tx

nTx

t

(n + 1)Tx (n + 2)Tx y(t) y(nTy)

(b)

t

(m − 1) Ty (m +1) Ty mTy x(t) (c)

g(nTy)

mTy

v(nTy)

t

(m + 2)Ty

Figure 1.4 Illustration of timing relations for sampling rate increase by an integer factor I = 2. The approach in (a) requires one impulse response for the even-numbered and one for the odd-numbered output samples. The approach in (c) requires only one impulse response, obtained by interleaving the impulse responses in (a).

2

Decimation by a Factor D Let us assume that the signal x(n) with spectrum X(ω) is to be downsampled by an integer factor D . The spectrum X(ω) is assumed to be nonzero in the frequency interval 0 ≤ |ω| ≤ π or, equivalently, |F | ≤ Fx /2. We know that if we reduce the sampling rate simply by selecting every Dth value of x(n), the resulting signal will be an aliased version of x(n), with a folding frequency of Fx /2D. To avoid aliasing, we must first reduce the bandwidth of x(n) to Fmax = Fx /2D or, equivalently, to ωmax = π/D . Then we may downsample by D and thus avoid aliasing. The decimation process is illustrated in Fig. 2.1. The input sequence x(n) is passed through a lowpass filter, characterized by the impulse response h(n) and a frequency response HD (ω), which ideally satisfies the condition  HD (ω) =

1, 0,

|ω| ≤ π/D otherwise

(2.1)

Thus the filter eliminates the spectrum of X(ω) in the range π/D < ω < π . Of course, the implication is that only the frequency components of x(n) in the range |ω| ≤ π/D are of interest in further processing of the signal.

773

Multirate Digital Signal Processing

x(n) Fx =

h(n)

(n)

y(m)

Downsampler ↓D

1 Tx

Fy =

Fx D

Figure 2.1 Decimation by a factor D .

The output of the filter is a sequence v(n) given as v(n) =

∞ 

h(k)x(n − k)

(2.2)

k=0

which is then downsampled by the factor D to produce y(m). Thus y(m) = v(mD) =

∞ 

h(k)x(mD − k)

(2.3)

k=0

Although the filtering operation on x(n) is linear and time invariant, the downsampling operation in combination with the filtering results in a time-variant system. This is easily verified. Given the fact that x(n) produces y(m), we note that x(n − n0 ) does not imply y(n−n0 ) unless n0 is a multiple of D. Consequently, the overall linear operation (linear filtering followed by downsampling) on x(n) is not time invariant. The frequency-domain characteristics of the output sequence y(m) can be obtained by relating the spectrum of y(m) to the spectrum of the input sequence x(n). First, it is convenient to define a sequence v˜ (n) as  v˜ (n) =

v(n), 0,

n = 0, ±D, ±2D, . . . otherwise

(2.4)

Clearly, v˜ (n) can be viewed as a sequence obtained by multiplying v(n) with a periodic train of impulses p(n), with period D , as illustrated in Fig. 2.2. The discrete Fourier series representation of p(n) is

p(n) =

D−1 1  j 2πkn/D e D

(2.5)

k=0

Hence v˜ (n) = v(n)p(n)

(2.6)

y(m) = v˜ (mD) = v(mD)p(mD) = v(mD)

(2.7)

and

774

Multirate Digital Signal Processing

v(n) 6 0

3

0

3

0

3

n

p(n)

6

n

~ v(n) = v(n)p(n) 6

Figure 2.2

Steps required to facilitate the mathematical description of downsampling by a factor D , using a sinusoidal sequence for illustration.

n

y(n) = ~ v(nD) = v(nD) 3 0

6

n

Now the z-transform of the output sequence y(m) is Y (z) =

∞ 

y(m)z−m

m=−∞

=

∞ 

v˜ (mD)z−m

(2.8)

m=−∞

Y (z) =

∞ 

v˜ (m)z−m/D

m=−∞

where the last step follows from the fact that v˜ (m) = 0, except at multiples of D . By making use of the relations in (2.5) and (2.6) in (2.8), we obtain  D−1  ∞  1  j 2πmk/D −m/D Y (z) = v(m) e z D m=−∞ k=0

D−1 ∞ 1   = v(m)(e−j 2πk/D z1/D )−m D m=−∞ k=0

D−1 1  = V (e−j 2πk/D z1/D ) D

(2.9)

k=0

=

D−1 1  HD (e−j 2πk/D z1/D )X(e−j 2πk/D z1/D ) D k=0

where the last step follows from the fact that V (z) = HD (z)X(z).

775

Multirate Digital Signal Processing

By evaluating Y (z) in the unit circle, we obtain the spectrum of the output signal y(m). Since the rate of y(m) is Fy = 1/Ty , the frequency variable, which we denote as ωy , is in radians and is relative to the sampling rate Fy , ωy =

2π F = 2π F Ty Fy

(2.10)

Since the sampling rates are related by the expression Fy =

Fx D

(2.11)

it follows that the frequency variables ωy and ωx =

2π F = 2π F Tx Fx

(2.12)

ωy = Dωx

(2.13)

are related by Thus, as expected, the frequency range 0 ≤ |ωx | ≤ π/D is stretched into the corresponding frequency range 0 ≤ |ωy | ≤ π by the downsampling process. We conclude that the spectrum Y (ωy ), which is obtained by evaluating (2.9) on the unit circle, can be expressed as     D−1 ωy − 2π k ωy − 2π k 1  Y (ωy ) = HD X D D D

(2.14)

k=0

With a properly designed filter HD (ω), the aliasing is eliminated and, consequently, all but the first term in (2.14) vanish. Hence ω ω

1 y y HD X D D D 1 ωy

= X D D

Y (ωy ) =

(2.15)

for 0 ≤ |ωy | ≤ π . The spectra for the sequences x(n), v(n), and y(m) are illustrated in Fig. 2.3. EXAMPLE 2.1 Design a decimator that downsamples an input signal x(n) by a factor D = 2. Use the Remez algorithm to determine the coefficients of the FIR filter that has a 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Solution. A filter length M = 30 achieves the design specifications given above. The frequency response is illustrated in Fig. 2.4. Note that the cutoff frequency is ωc = π/2.

776

Multirate Digital Signal Processing

X(ωx )

−π

π ωx

0 HD(ωx )

−π

−π



π D

0



π D

0

π D V(ωx )

ωx

π D Y(ωy)

ωx

Figure 2.3

Spectra of signals in the decimation of x(n) by a factor D .

−π

0

π

ωy

Figure 2.4 Magnitude response of linear-phase FIR filter of length M = 30 in Example 2.1.

777

Multirate Digital Signal Processing

3

Interpolation by a Factor I An increase in the sampling rate by an integer factor of I can be accomplished by interpolating I − 1 new samples between successive values of the signal. The interpolation process can be accomplished in a variety of ways. We shall describe a process that preserves the spectral shape of the signal sequence x(n). Let v(m) denote a sequence with a rate Fy = I Fx , which is obtained from x(n) by adding I − 1 zeros between successive values of x(n). Thus  x(m/I ), m = 0, ±I, ±2I, . . . (3.1) v(m) = 0, otherwise and its sampling rate is identical to the rate of y(m). This sequence has a z-transform V (z) =

∞ 

v(m)z−m

m=−∞

=

∞ 

x(m)z−mI

(3.2)

m=−∞

= X(zI ) The corresponding spectrum of v(m) is obtained by evaluating (3.2) on the unit circle. Thus V (ωy ) = X(ωy I ) (3.3) where ωy denotes the frequency variable relative to the new sampling rate Fy (i.e., ωy = 2πF /Fy ). Now the relationship between sampling rates is Fy = I Fx and hence, the frequency variables ωx and ωy are related according to the formula ωy =

ωx I

(3.4)

The spectra X(ωx ) and V (ωy ) are illustrated in Fig. 3.1. We observe that the sampling rate increase, obtained by the addition of I − 1 zero samples between successive values of x(n), results in a signal whose spectrum V (ωy ) is an I -fold periodic repetition of the input signal spectrum X(ωx ). Since only the frequency components of x(n) in the range 0 ≤ ωy ≤ π/I are unique, the images of X(ω) above ωy = π/I should be rejected by passing the sequence v(m) through a lowpass filter with frequency response HI (ωy ) that ideally has the characteristic  C, 0 ≤ |ωy | ≤ π/I HI (ωy ) = (3.5) 0, otherwise where C is a scale factor required to properly normalize the output sequence y(m). Consequently, the output spectrum is  CX(ωy I ), 0 ≤ |ωy | ≤ π/I Y (ωy ) = (3.6) 0, otherwise

778

Multirate Digital Signal Processing

|X(ωx)|

−π

π

0

ωx

|V(ωy)| I=2



2π I



π I

0

π I

2π I

ωy

π

ωy

π

ωy

|HI(ωy)|

−π



π I

0

π I |Y(ωy)|

Figure 3.1

Spectra of x(n) and v(n) where V (ωy ) = X(ωy l ).

−π



π I

0

π I

The scale factor C is selected so that the output y(m) = x(m/I ) for m = 0, ±I , +2I, . . . . For mathematical convenience, we select the point m = 0. Thus π 1 y(0) = Y (ωy )dωy 2π −π (3.7) π/I C = X(ωy I )dωy 2π −π/I Since ωy = ωx /I , (3.7) can be expressed as π C 1 X(ωx )dωx y(0) = I 2π −π

(3.8)

C = x(0) I Therefore, C = I is the desired normalization factor. Finally, we indicate that the output sequence y(m) can be expressed as a convolution of the sequence v(n) with the unit sample response h(n) of the lowpass filter. Thus ∞  y(m) = h(m − k)v(k) (3.9) k=−∞

779

Multirate Digital Signal Processing

Figure 3.2 Magnitude response of linear-phase FIR filter of length M =

30 in Example 3.1.

Since v(k) = 0 except at multiples of I , where v(kI ) = x(k), (3.9) becomes y(m) =

∞ 

h(m − kI )x(k)

(3.10)

k=−∞

EXAMPLE 3.1 Design a interpolator that increases the input sampling rate by a factor of l = 5. Use the Remez algorithm to determine the coefficients of the FIR filter that has a 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Solution. A filter length M = 30 achieves the design specifications given above. The frequency response of the FIR filter is illustrated in Fig. 3.2. Note that the cutoff frequency is ωc = π/5.

4

Sampling Rate Conversion by a Rational Factor I / D Having discussed the special cases of decimation (downsampling by a factor D) and interpolation (upsampling by a factor I ), we now consider the general case of sampling rate conversion by a rational factor I /D. Basically, we can achieve this sampling rate conversion by first performing interpolation by the factor I and then decimating the output of the interpolator by the factor D. In other words, a sampling rate conversion by the rational factor I /D is accomplished by cascading an interpolator with a decimator, as illustrated in Fig. 4.1. We emphasize that the importance of performing the interpolation first and the decimation second is to preserve the desired spectral characteristics of x(n). Furthermore, with the cascade configuration illustrated in Fig. 4.1, the two filters with impulse response {hu (k)} and {hd (k)} are operated at the same rate, namely I Fx , and hence can be combined into a single lowpass filter with impulse response h(k) as illustrated in Fig. 4.2. The frequency response H (ωv ) of the combined filter

780

Multirate Digital Signal Processing

x(n) Rate Fx

Upsampler ↑I

Filter hu(k)

Interpolator

y(m)

Downsampler ↓D

Filter hd(k)

Decimator

Rate = I Fx

Rate =

I Fx = Fy D

Figure 4.1 Method for sampling rate conversion by a factor I /D .

must incorporate the filtering operations for both interpolation and decimation, and hence it should ideally possess the frequency response characteristic  H (ωv ) =

0 ≤ |ωv | ≤ min(π/D, π/I ) otherwise

I, 0,

(4.1)

where ωv = 2πF /Fv = 2π F /I Fx = ωx /I . In the time domain, the output of the upsampler is the sequence  v(l) =

x(l/I ), 0,

l = 0, ±I, ±2I, . . . otherwise

(4.2)

and the output of the linear time-invariant filter is w(l) =

∞ 

h(l − k)v(k)

k=−∞

=

∞ 

(4.3) h(l − kI )x(k)

k=−∞

Finally, the output of the sampling rate converter is the sequence {y(m)}, which is

x(n) Rate = Fx

Upsampler ↑I

ν(k)

Lowpass filter h(k)

w(l)

Downsampler ↓D

y(m) Rate =

I Fx = F y D

Rate = IFx = Fν

Figure 4.2 Method for sampling rate conversion by a factor I /D .

781

Multirate Digital Signal Processing

obtained by downsampling the sequence {w(l)} by a factor of D . Thus y(m) = w(mD) =

∞ 

h(mD − kI )x(k)

(4.4)

k=−∞

It is illuminating to express (4.4) in a different form by making a change in variable. Let   mD −n k= (4.5) I where the notation r denotes the largest integer contained in r . With this change in variable, (4.4) becomes y(m) =

       mD mD h mD − I + nI x −n I I n=−∞ ∞ 

We note that

 mD −

 mD I = mD, I

(4.6)

modulo I

= (mD)I Consequently, (4.6) can be expressed as y(m) =

∞ 

 h(nI + (mD)I )x

n=−∞

  mD −n I

(4.7)

which is the discrete-time version of (1.9). It is apparent from this form that the output y(m) is obtained by passing the input sequence x(n) through a time-variant filter with impulse response g(n, m) = h(nI + (mD)I ),

−∞ < m, n < ∞

(4.8)

where h(k) is the impulse response of the time-invariant lowpass filter operating at the sampling rate I Fx . We further observe that for any integer k, g(n, m + kI ) = h(nI + (mD + kDI )I ) = h(nI + (mD)I ) = g(n, m) Hence g(n, m) is periodic in the variable m with period I .

782

(4.9)

Multirate Digital Signal Processing

The frequency-domain relationships can be obtained by combining the results of the interpolation and decimation processes. Thus the spectrum at the output of the linear filter with impulse response h(l) is V (ωv ) = H (ωv )X(ωv I )  I X(ωv I ), 0 ≤ |ωv | ≤ min(π/D, π/I ) = 0, otherwise

(4.10)

The spectrum of the output sequence y(m), obtained by decimating the sequence v(n) by a factor of D, is Y (ωy ) =

  D−1 ωy − 2π k 1  V D D

(4.11)

k=0

where ωy = Dωv . Since the linear filter prevents aliasing as implied by (4.10), the spectrum of the output sequence given by (4.11) reduces to    πD  I ωy

X | ≤ min π, , 0 ≤ |ω y Y (ωy ) = D (4.12) D I  0, otherwise EXAMPLE 4.1 Design a sample rate converter that increases the sampling rate by a factor of 2.5. Use the Remez algorithm to determine the coefficients of the FIR filter that has a 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Specify the sets of time-varying coefficients g(n, m) used in the realization of the sampling rate converter. Solution. The FIR filter that meets the specifications of this problem is exactly the same as the filter designed in Example 3.1. Its bandwidth is π/5. The coefficients of the FIR filter are given by (4.8) as g(n, m) = h(nI + (mD)I )   mD I = h nI + mD −  I By substituting I = 5 and D = 2 we obtain   2m  g(n, m) = h 5n + 2m − 5 5 By evaluating g(n, m) for n = 0, 1, . . . , 5 and m = 0, 1, . . . , 4 we obtain the following coefficients for the time-variant filter: g(0, m) = {h(0) g(1, m) = {h(5) g(2, m) = {h(10) g(3, m) = {h(15) g(4, m) = {h(20) g(0, m) = {h(25)

h(2) h(7) h(12) h(17) h(22) h(27)

h(4) h(9) h(14) h(19) h(24) h(29)

h(1) h(6) h(11) h(16) h(21) h(26)

h(3) } h(8) } h(13)} h(18)} h(23)} h(28)}

783

Multirate Digital Signal Processing

In summary, sampling rate conversion by the factor I /D can be achieved by first increasing the sampling rate by I , accomplished by inserting I − 1 zeros between successive values of the input signal x(n), followed by linear filtering of the resulting sequence to eliminate unwanted images of X(ω) and, finally, by downsampling the filtered signal by the factor D. When Fy > Fx , the lowpass filter acts as an anti-imaging postfilter that removes the spectral replicas at multiples of Fx , but not at multiples of I Fx . When Fy < Fx , the lowpass filter acts as an anti-aliasing prefilter that removes the down-shifted spectral replicas at multiples of Fy to avoid overlapping.

5

Implementation of Sampling Rate Conversion In this section we consider the efficient implementation of sampling rate conversion systems using polyphase filter structures. Additional computational simplifications can be obtained using the multistage approach described in Section 6.

5.1

Polyphase Filter Structures

Polyphase structures for FIR filters were developed for the efficient implementation of sampling rate converters; however, they can be used in other applications. The polyphase structure is based on the fact that any system function can be split as + h(M)z−M + · · ·

H (z) = · · · + h(0) · · · + h(1)z−1

+ h(M + 1)z−(M+1) + · · · .. .

· · · + h(M − 1)z−(M−1) + h(2M − 1)z−(2M−1) + · · · If we next factor out the term z−(i−1) at the i th row, we obtain + h(M)z−M + · · ·]

H (z) = [· · · + h(0) + z−1 [· · · + h(1)

+ h(M + 1)z−M + · · ·] .. .

+ z−(M−1) [· · · + h(M − 1) + h(2M − 1)z−M + · · ·] The last equation can be compactly expressed as H (z) =

M−1 

z−i Pi (zM )

(5.1)

i=0

where Pi (z) =

∞  n=−∞

784

h(nM + i)z−n

(5.2)

Multirate Digital Signal Processing

P0(z3) x(n)

y(n) z −1 P1(z3) z −1

P2(z3) Figure 5.1 Block diagram of polyphase filter structure for M = 3.

Relation (5.1) is called the M-component polyphase decomposition and P (z)i the polyphase components of H (z). Each subsequence pi (n) = h(nM + i),

i = 0, 1, . . . , M − 1

(5.3)

is obtained by downsampling a delayed (“phase-shifted”) version of the original impulse response. To develop an M -component polyphase filter structure, we use (5.1) for M = 3 to express the z-transform of the output sequence as Y (z) = H (z)X(z) = P0 (z3 )X(z) + z−1 P1 (z3 )X(z) + z−2 P2 (z3 )X(z)

(5.4)

= P0 (z3 )X(z) + z−1 {P1 (z3 )X(z) + z−1 [P2 (z3 )X(z)]}

(5.5)

Equation (5.4) leads to the polyphase structure of Figure 5.1. Similarly, (5.5) yields the polyphase structure shown in Figure 5.2. This is known as transpose polyphase structure because it is similar to the transpose FIR filter realization. The obtained polyphase structures are valid for any filter, FIR or IIR, and any finite value of M and are sufficient for our needs. Additional structures and details can be found in Vaidyanathan (1993).

5.2

Interchange of Filters and Downsamplers/Upsamplers

In general, the order of a sampling rate converter (which is a linear time-varying system) and a linear time-invariant system cannot be interchanged. We next derive two identities, known as noble identities, that help to swap the position of a filter with a downsampler or an upsampler by properly modifying the filter.

785

Multirate Digital Signal Processing

P0(z3) x(n)

y(n) z −1

P1(z3)

z −1

P2(z3) Figure 5.2 Illustration of transpose polyphase filter structure for M = 3.

To prove the first identity (see Figure 5.3), we recall that the input/output description of a downsampler is D−1 1  X(z1/D WDi ) y(n) = x(nD) ←→ Y (z) = D Z

(5.6)

i=0

where WD = e−j 2π/D . The output of the system in Figure 5.3(a) can be written as Y (z) =

D−1 D−1 1  1  V1 (z1/D WDi ) = H (zWDiD )X(z1/D WDi ) D D i=0

(5.7)

i=0

because V1 (z) = H (zD )X(z). Taking into consideration WDiD = 1 and Figure 5.3(b), relation (5.7) results in  1 X(z1/D WDi ) = H (z)V2 (z) H (z) D D−1

Y (z) =

(5.8)

i=0

which shows the equivalence of the two structures in Figure 5.3. H(z)

D v1(m)

x(n)

y(m)

(a)

Figure 5.3

Two equivalent downsampling systems (first noble identity).

786

D

H(zD) x(n)

v2(n) (b)

y(m)

Multirate Digital Signal Processing

H(z)

I v1(n)

x(n)

y(m)

(a)

I

Figure 5.4

Two equivalent upsampling systems (second noble identity).

H(zI) v2(m)

x(n)

y(m)

(b)

A similar identity can be shown to hold for upsampling. To start, we recall that the input/output description of an upsampler is y(n) =

   x nI , 0,

Z n = 0, ±I, ±2I, . . . ←→ Y (z) = X(zI ) otherwise

(5.9)

The output of the system in Figure 5.4(a) can be written as Y (z) = H (zI )V1 (z) = H (zI )X(zI )

(5.10)

because V1 (z) = X(zI ). The output of the system in Figure 5.4(b) is given by Y (z) = V2 (zI ) = H (zI )X(zI )

(5.11)

which is equal to (5.10). This shows that the two systems in Figure 5.4 are identical. In conclusion, we have shown that it is possible to interchange the operation of linear filtering and downsampling or upsampling if we properly modify the system function of the filter.

5.3

Sampling Rate Conversion with Cascaded Integrator Comb Filters

The hardware implementation of the lowpass filter required for sampling rate conversion can be significantly simplified if we choose a comb filter with system function H (z) =

M−1  k=0

z−k =

1 − z−M 1 − z−1

(5.12)

This system can be implemented by cascading either the “integrator” 1/(1−z−1 ) with the comb filter (1 − z−M ) or vice versa. This leads to the name cascaded integrator comb (CIC) filter structure. The CIC structure does not require any multiplications or storage for the filter coefficients. To obtain an efficient decimation structure, we start with an integrator-comb CIC filter followed by the downsampler and then we apply the first noble identity, as shown in Figure 5.5. For the interpolation case, we use an upsampler followed

787

Multirate Digital Signal Processing

x(n)

1 1 − z−1

1 − z−D

D y(m)

(a)

x(n)

1 1 − z−1

D

1 − z−1 y(m)

(b) Figure 5.5 Using the first noble identity to obtain an efficient CIC filter structure for decimation.

by a comb-integrator CIC filter and then we use the second noble identity, as shown in Figure 5.6. To improve the lowpass frequency response required for sampling rate conversion, we can cascade K CIC filters. In this case, we order all integrators on one side of the filter and the comb filters on the other side, and then we apply the noble identities as in the single-stage case. The integrator 1/(1 − z−1 ) is an unstable system. Therefore, its output may grow without limits, resulting in overflow when the integrator section comes first, as in the decimation structure shown in Figure 5.5(b). However, this overflow can be tolerated if the entire filter is implemented using two’s-complement fixed-point arithmetic. If D = M or I = M , the comb filter 1−z−1 in Figures 5.5(a) and 5.6(b) should be replaced by 1−z −M/D or 1 − z−M/I , respectively. A detailed treatment of CIC filters for decimation and interpolation can be found in Hogenauer (1981). Finally, we note that CIC filters are special cases of frequency sampling structure. If the CIC filter order is a power of 2, that is, M = 2K , we can decompose the system function (5.12) as follows: H (z) = (1 + z−1 )(1 + z−2 )(1 + z−4 ) . . . (1 + z−2

K−1

(5.13)

)

Using this decomposition we can develop decimator structures using nonrecursive

I

1 − z−I

x(n)

1 1 − z−1

y(m)

1 1 − z−1

y(m)

(a)

1 − z−1

I

x(n) (b)

Figure 5.6 Using the second noble identity to obtain an efficient CIC

filter structure for interpolation.

788

Multirate Digital Signal Processing

2

2

x(n)

2 y(m) z −1

z −1

z −1

Figure 5.7 Efficient filter structure for decimation by D = 8 using comb filters.

CIC filters. Figure 5.7 shows an example for a decimator with D = M = 8. Cascading of N CIC filters can be obtained by providing M first-order sections (1 − z−1 ) between each decimation stage. The constraint M = 2K can be relaxed by factoring M into a product of prime numbers, as shown in Jang and Yang (2001).

5.4

Polyphase Structures for Decimation and Interpolation Filters

To develop a polyphase structure for decimation, we start with the straightforward implementation of the decimation process shown in Figure 5.8. The decimated sequence is obtained by passing the input sequence x(n) through a linear filter and then downsampling the filter output by a factor D. In this configuration, the filter is operating at the high sampling rate Fx , while only one out of every D output samples is actually needed. A logical solution would be to find a structure where only the needed samples are computed. We will develop such an efficient implementation by exploitating the polyphase structure in Figure 5.1. Since downsampling commutes with addition, combining the structures in Figures 5.8 and 5.1 yields the structure in Figure 5.9(a). If we next apply the identity in Figure 5.3, we obtain the desired implementation structure shown in Figure 5.9(b). In this filtering structure, only the needed samples are computed and all multiplications and additions are performed at the lower sampling rate Fx /D . Thus, we have achieved the desired efficiency. Additional reduction in computation can be achieved by using an FIR filter with linear phase and exploiting the symmetry of its impulse response. In practice it is more convenient to implement the polyphase decimator using a commutator model as shown in Figure 5.10. The commutator rotates counterclockwise starting at time n = 0 and distributes a block of D input samples to the polyphase filters starting at filter i = D − 1 and continuing in reverse order until i = 0. For every block of D input samples, the polyphase filters receive a new input and their outputs are computed and summed to produce one sample of the output signal y(m). The operation of this realization can be also understood by a careful inspection of Figure 1.3. Next, let us consider the efficient implementation of an interpolator, which is realized by first inserting I − 1 zeros between successive samples of x(n) and then

D

H(z) x(n)

v(n)

y(m) = v(mD)

Figure 5.8 Decimation system.

789

Multirate Digital Signal Processing

P0(z3)

x(n)

3 y(m)

z −1 P1(z3)

3

P2(z3)

3

z −1

(a) 3

P0(z)

x(n)

y(m) z −1 3

P1(z)

3

P2(z)

z −1

(b) Figure 5.9 Implementation of a decimation system using a polyphase structure before (a) and after (b) the use of the first noble identity.

filtering the resulting sequence (see Figure 5.11). The major problem with this structure is that the filter computations are performed at the high sampling rate I Fx . The desired simplification is achieved by first replacing the filter in Figure 5.11 with the transpose polyphase structure in Figure 5.2, as illustrated in Figure 5.12(a). Fy = Fx / D Commutator rate = Fx

Fx x(n)

P0(z) y(m) P1(z)

.. .

.. .

PD −1(z) Figure 5.10 Decimation using a polyphase filter and a commutator.

790

Multirate Digital Signal Processing

H(z)

I

y(m)

v(m)

x(n)

Figure 5.11 Interpolation system.

P0(z3)

3 x(n)

y(m) z −1

3

P1(z3)

z −1

3

P2(z3)

(a)

P0(z)

3

x(n)

y(m) z −1

P1(z)

3

z −1

P2(z)

3

(b) Figure 5.12 Implementation of an interpolation system using a polyphase

structure before (a) and after (b) the use of the second noble identity.

791

Multirate Digital Signal Processing

Then, we use the second noble identity (see Figure 5.4) to obtain the structure in Figure 5.12 (b). Thus, all the filtering multiplications are performed at the low rate Fx . It is interesting to note that the structure for an interpolator can be obtained by transposing the structure of a decimator, and vice versa (Crochiere and Rabiner, 1981). For every input sample, the polyphase filters produce I output samples y0 (n), y1 (n), . . ., yI −1 (n). Because the output yi (n) of the i th filter is followed by (I − 1) zeros and it is delayed by i th samples, the polyphase filters contribute nonzero samples at different time slots. In practice, we can implement the part of the structure including the 1-to-I expanders, delays, and adders using the commutator model shown in Figure 5.13. The commutator rotates counterclockwise starting at time n = 0 at branch i = 0. For each input sample x(n) the commutator reads the output of every polyphase filter to obtain I samples of the output (interpolated) signal y(m). The operation of this realization can be also understood by a careful inspection of Figure 1.4. Each polyphase filter in Figure 5.13 operates on the same input data using its unique set of coefficients. Therefore, we can obtain the same results using a single filter by sequentially loading a different set of coefficients.

5.5

Structures for Rational Sampling Rate Conversion

A sampling rate converter with a ratio I /D can be efficiently implemented using a polyphase interpolator followed by a downsampler. However, since the downsampler keeps only every Dth polyphase subfilter output, it is not necessary to compute all I interpolated values between successive input samples. To determine which polyphase subfilter outputs to compute, we consider an example with I = 5 and D = 3. The interpolator polyphase structure has I = 5 subfilters which provide interpolated samples at an effective sampling period T = Tx /I . The downsampler picks every Dth of these samples, resulting in a discrete-time signal with sampling Fx x(n)

P0(z) Commutator rate = I Fx P1(z)

.. .

Fy = I Fx y(m)

PI−1(z) Figure 5.13 Interpolation using a polyphase filter and a commutator.

792

Multirate Digital Signal Processing

Tblock = 3Tx = 5Ty = 15T

Tx

T Input Interpolated Output

Ty 1

2

3

4

t

n

0

k

0 1 2 3 4 5 6 7 8 9 10

...

15

m

0

1

2

3

4

5

6

7

8

9

km

0

0

1

1

2

3

3

4

4

5

im

0

3

1

4

2

0

3

1

4

2

...

5

20

...

25

Figure 5.14 Illustration of index computation for polyphase implementation of

sampling rate conversion for a rational ratio I /D = 5/3.

period Ty = DT = DTx /I . It is convenient to think in terms of blocks of duration Tblock = I Ty = DTx = I DT

(5.14)

which contain L output samples or I input samples. The relative time positions of the various sequences and a block of data are illustrated in Figure 5.14. The input sequence x(nTx ) is interpolated to produce a sequence v(kT ), which is then decimated to yield y(mTy ). If we use an FIR filter with M = KI coefficients, the polyphase subfilters are given by pi (n) = h(nI + i), i = 0, 1, . . . , I − 1, where n = 0, 1, . . . , K − 1. To compute the output sample y(m), we use the polyphase subfilter with index im which requires the input samples x(km ), x(km − 1), . . . , x(km − K + 1). From relation (1.9) and Figure 5.14, we can easily deduce that  km =

mD I

 and im = (Dm)I

(5.15)

For D = 3 and I = 5 the first data block includes D = 3 input samples and I = 5 output samples. To compute the samples {y(0), y(1), y(2), y(3), y(4)}, we use the polyphase subfilter specified by the index im = {0, 3, 1, 4, 2}, respectively. The samples in the filter memory are updated only when km changes value. This discussion provides the basic ideas for the efficient software implementation of rational sampling rate conversion using FIR filters.

6

Multistage Implementation of Sampling Rate Conversion In practical applications of sampling-rate conversion we often encounter decimation factors and interpolation factors that are much larger than unity. For example, suppose that we are given the task of altering the sampling rate by the factor

793

Multirate Digital Signal Processing

I1Fx x(n) Fx

↑I1

↑I2

h1(n) Stage 1



h2(n)

↑IL

I1I2Fx

Stage 2

hL(n)

y(m) Fy = IFx

Stage L

Figure 6.1 Multistage implementation of interpolation by a factor I .

I /D = 130/63. Although, in theory, this rate alteration can be achieved exactly, the implementation would require a bank of 130 polyphase filters and may be computationally inefficient. In this section we consider methods for performing sampling rate conversion for either D 1 and/or I 1 in multiple stages. First, let us consider interpolation by a factor I 1 and let us assume that I can be factored into a product of positive integers as I=

L 

(6.1)

Ii

i=1

Then, interpolation by a factor I can be accomplished by cascading L stages of interpolation and filtering, as shown in Fig. 6.1. Note that the filter in each of the interpolators eliminates the images introduced by the upsampling process in the corresponding interpolator. In a similar manner, decimation by a factor D, where D may be factored into a product of positive integers as D=

J 

(6.2)

Di

i=1

can be implemented as a cascade of J stages of filtering and decimation as illustrated in Fig. 6.2. Thus the sampling rate at the output of the i th stage is Fi−1 , Di

Fi =

i = 1, 2, . . . , J

(6.3)

where the input rate for the sequence {x(n)} is F0 = Fx . Fx D1 x(n) Fx

h1(n) Stage 1

↓D1

h2(n) Stage 2

↓D2

… Fx D1D2

hJ (n) Stage J

Figure 6.2 Multistage implementation of decimation by a factor D .

794

↓DJ

y(m) Fx D

Multirate Digital Signal Processing

To ensure that no aliasing occurs in the overall decimation process, we can design each filter stage to avoid aliasing within the frequency band of interest. To elaborate, let us define the desired passband and the transition band in the overall decimator as Passband: 0 ≤ F ≤ Fpc Transition band: Fpc ≤ F ≤ Fsc

(6.4)

where Fsc ≤ Fx /2D. Then, aliasing in the band 0 ≤ F ≤ Fsc is avoided by selecting the frequency bands of each filter stage as follows: Passband: 0 ≤ F ≤ Fpc Transition band: Fpc ≤ F ≤ Fi − Fsc

(6.5)

Fi−1 Stopband: Fi − Fsc ≤ F ≤ 2 For example, in the first filter stage we have F1 = Fx /D1 , and the filter is designed to have the following frequency bands: Passband: 0 ≤ F ≤ Fpc Transition band: Fpc ≤ F ≤ F1 − Fsc

(6.6)

F0 Stopband: F1 − Fsc ≤ F ≤ 2 After decimation by D1 , there is aliasing from the signal components that fall in the filter transition band, but the aliasing occurs at frequencies above Fsc . Thus there is no aliasing in the frequency band 0 ≤ F ≤ Fsc . By designing the filters in the subsequent stages to satisfy the specifications given in (6.5), we ensure that no aliasing occurs in the primary frequency band 0 ≤ F ≤ Fsc . EXAMPLE 6.1 Consider an audio-band signal with a nominal bandwidth of 4 kHz that has been sampled at a rate of 8 kHz. Suppose that we wish to isolate the frequency components below 80 Hz with a filter that has a passband 0 ≤ F ≤ 75 and a transition band 75 ≤ F ≤ 80. Hence Fpc = 75 Hz and Fsc = 80. The signal in the band 0 ≤ F ≤ 80 may be decimated by the factor D = Fx /2Fsc = 50. We also specify that the filter have a passband ripple δ1 = 10−2 and a stopband ripple of δ2 = 10−4 . The length of the linear phase FIR filter required to satisfy these specifications can be estimated from a well-known formula. Recall that a particularly simple formula for appr oximating the length M, attributed to Kaiser, is ˆ = −10 log10 δ1 δ2 − 13 + 1 M 14.6f

(6.7)

where f is the normalized (by the sampling rate) width of the transition region [i.e., f = (Fsc − Fpc )/Fs ]. A more accurate formula proposed by Herrmann et al. (1973) is 2 ˆ = D∞ (δ1 , δ2 ) − f (δ1 , δ2 )(f ) + 1 M f

(6.8)

795

Multirate Digital Signal Processing

where D∞ (δ1 , δ2 ) and f (δ1 , δ2 ) are defined as D∞ (δ1 , δ2 ) = [0.005309(log10 δ1 )2 + 0.07114(log10 δ1 ) − 0.4761] log10 δ2 − [0.00266(log10 δ1 )2 + 0.5941 log10 δ1 + 0.4278] f (δ1 , δ2 ) = 11.012 + 0.51244[log10 δ1 − log10 δ2 ]

(6.9) (6.10)

Now a single FIR filter followed by a decimator would require (using the Kaiser formula) a filter of (approximate) length −6 ˆ = −10 log10 10 − 13 + 1 ≈ 5152 M 14.6(5/8000)

As an alternative, let us consider a two-stage decimation process with D1 = 25 and D2 = 2. In the first stage we have the specifications F1 = 320 Hz and Passband: 0 ≤ F ≤ 75 Transition band: 75 < F ≤ 240 f = δ11 =

δ1 , 2

165 8000

δ21 = δ2

Note that we have reduced the passband ripple δ1 by a factor of 2, so that the total passband ripple in the cascade of the two filters does not exceed δ1 . On the other hand, the stopband ripple is maintained at δ2 in both stages. Now the Kaiser formula yields an estimate of M1 as ˆ 1 = −10 log10 δ11 δ21 − 13 + 1 ≈ 167 M 14.6f For the second stage, we have F2 = F1 /2 = 160 and the specifications Passband: 0 ≤ F ≤ 75 Transition band: 75 < F ≤ 80 f = δ12 =

δ1 , 2

5 320

δ22 = δ2

Hence the estimate of the length M2 of the second filter is ˆ 2 ≈ 220 M ˆ1 +M ˆ 2 = 387. This Therefore, the total length of the two FIR filters is approximately M represents a reduction in the filter length by a factor of more than 13. The reader is encouraged to repeat the computation above with D1 = 10 and D2 = 5.

796

Multirate Digital Signal Processing

It is apparent from the computations in Example 6.1 that the reduction in the filter length results from increasing the factor f , which appears in the denominator in (6.7) and (6.8). By decimating in multiple stages, we are able to increase the width of the transition region through a reduction in the sampling rate. In the case of a multistage interpolator, the sampling rate at the output of the i th stage is Fi−1 = Ii Fi , i = J, J − 1, . . . , 1 and the output rate is F0 = I FJ when the input sampling rate is FJ . The corresponding frequency band specifications are Passband: 0 ≤ F ≤ Fp Transition band: Fp < F ≤ Fi − Fsc The following example illustrates the advantages of multistage interpolation. EXAMPLE 6.2 Let us reverse the filtering problem described in Example 6.1 by beginning with a signal having a passband 0 ≤ F ≤ 75 and a transition band of 75 ≤ F ≤ 80. We wish to interpolate by a factor of 50. By selecting I1 = 2 and I2 = 25, we have basically a transposed form of the decimation problem considered in Example 6.1. Thus we can simply transpose the ˆ 1 ≈ 220, two-stage decimator to achieve the two-stage interpolator with I1 = 2, I2 = 25, M ˆ 2 ≈ 167. and M

7

Sampling Rate Conversion of Bandpass Signals In this section we consider the decimation and interpolation of bandpass signals. We begin by noting that any bandpass signal can be converted into an equivalent lowpass signal whose sampling rate can be changed using the already developed techniques. However, a simpler and more widely used approach concerns bandpass discrete-time signals with integer-band positioning. The concept is similar to one for continuous-time bandpass signals. To be specific, suppose that we wish to decimate by a factor D an integerpositioned bandpass signal with spectrum confined to the bands (k − 1)

π π < |ω| < k D D

(7.1)

where k is a positive integer. A bandpass filter defined by  HBP (ω) =

π π 1, (k − 1) D < |ω| < k D 0, otherwise

(7.2)

would normally be used to eliminate signal frequency components outside the desired frequency range. Then direct decimation of the filtered signal v(n) by the factor D

797

Multirate Digital Signal Processing

results in a periodic replication of the bandpass spectrum V (ω) every 2π/D radians according to (2.14). The spectrum of the decimated signal y(m) is obtained by scaling the frequency axis by ωy = Dωx . This process is illustrated in Figure 7.1 for bandpass signal with odd band positioning (k = 3) and in Figure 7.2 for signals with even band positioning (k = 4). In the case where k is odd, there is an inversion of the spectrum of the signal as in the continuous-time case. The inversion can be undone by the simple process y (m) = (−1)m y(m). Note that violation of the bandwidth constraint given by (7.1) results in signal aliasing. The process of bandpass interpolation by an integer factor I is the inverse of that of bandpass decimation and can be accomplished in a similar manner. The process of upsampling by inserting zeros between the samples of x(n) produces I images in the band 0 ≤ ω ≤ π . The desired image can be selected by bandpass filtering. This can be seen by “reversing” the process shown in Figure 7.1. Note that the process of interpolation also provides us with the opportunity to achieve frequency translation of the spectrum. Finally, sampling rate conversion for a bandpass signal by a rational factor I /D can be accomplished by cascading a decimator with an interpolator in a manner that depends on the choice of the parameters D and I . A bandpass filter preceding the sampling converter is usually required to isolate the signal frequency band of

HBP(z)

x(n)

D v(n)

y(m) = v(mD)

(a) D=4 V(ω)

−3π/D

−2π/D

2π/D

0

ωx

3π/D

(b) D=4 Y(ω)

−3π/D

0 (c)

−2π/D

2π/D

ωx

3π/D

D=4 Y(ω)

−2π

−π

0 (d)

π



ωy

Figure 7.1 Spectral interpretation of bandpass decimation for integerband positioning (odd integer positioning).

798

Multirate Digital Signal Processing

V(ω)

−4π/D

−3π/D

3π/D

0

4π/D

ωx

(a) D=4 Y(ω)

−2π

−π

0 (b)

π



ωy

Figure 7.2 Spectral interpretation of bandpass decimation for integerband positioning (even integer positioning).

interest. Note that this approach provides us with a modulation-free method for achieving frequency translation of a signal by selecting D = I .

8

Sampling Rate Conversion by an Arbitrary Factor Efficient implementation of sampling rate conversion by a polyphase structure requires that the rates Fx and Fy are fixed and related by a rational factor I /D. In some applications, it is either inefficient or sometimes impossible to use such an exact rate conversion scheme. For example, suppose we need to perform rate conversion by the rational number I /D, where I is a large integer (e.g., I /D = 1023/511). Although we can achieve exact rate conversion by this number, we would need a polyphase filter with 1023 subfilters. Such an implementation is obviously inefficient in memory usage because we need to store a large number of filter coefficients. In some applications, the exact conversion rate is not known when we design the rate converter, or the rate is continuously changing during the conversion process. For example, we may encounter the situation where the input and output samples are controlled by two independent clocks. Even though it is still possible to define a nominal conversion rate that is a rational number, the actual rate would be slightly different, depending on the frequency difference between the two clocks. Obviously, it is not possible to design an exact converter in this case. In principle, we can convert from any rate Fx to any rate Fy (fixed or variable) using formula (1.9), which we repeat here for convenience: y(mTy ) =

K2 

g(kTx + m Tx )x((km − k)Tx )

(8.1)

k=K1

This requires the computation of a new impulse response pm (k) = g(kTx + m Tx ) for each output sample. However, if m is measured with finite accuracy, there is only a finite set of impulse responses, which may be precomputed and loaded from memory

799

Multirate Digital Signal Processing

as needed. We next discuss two practical approaches for sampling rate conversion by an arbitrary factor.

8.1

Arbitrary Resampling with Polyphase Interpolators

If we use a polyphase interpolator with I subfilters we can generate samples with spacing Tx /I . Therefore, the number I of stages determines the granularity of the interpolating process. If Tx /I is sufficiently small so that successive values of the signal do not change significantly or the change is less than the quantization step, we can determine the value at any location t = nTx + Tx , 0 ≤  ≤ 1, using the value of the nearest neighbor (zero-order hold interpolation). Additional improvement can be obtained using two-point linear interpolation y(nTx + Tx ) = (1 − )x(n) + x(n + 1)

(8.2)

The performance of these interpolation techniques by analyzing their frequencydomain characteristics. Additional practical details can be found in Ramstad (1984).

8.2

Arbitrary Resampling with Farrow Filter Structures

In practice, we typically implement polyphase sampling rate converters by a rational factor using causal FIR lowpass filters. If we use an FIR filter with M = KI coefficients, the coefficients of the polyphase filters are obtained by the mapping pi (n) = h(nI + i),

i = 0, 1, . . . , I − 1

(8.3)

This mapping can be easily visualized as mapping the one-dimensional sequence h(n) into a two-dimensional array with I rows and K columns by filling successive columns in natural order as follows: p0 (k) p1 (k)

→ h(0) → h(1) .. .

h(I ) h(I + 1)

... ...

h((K − 1)I ) h((K − 1)I + 1)

pi (k) pi+1 (k)

→ h(i) → h(i + 1) .. .

h(I + i) ... h(I + i + 1) . . .

h((K − 1)I + i)

pI −1 (k) → h(I − 1) h(2I − 1)

...

(8.4)

h(KI − 1)

The polyphase filters pi (n) are used to compute samples at I equidistant locations t = nTx + i(Tx /I ), i = 0, 1, . . . , I − 1 covering each input sampling interval. Suppose now that we wish to compute a sample at t = nTx + Tx , where  = i/I and 0 ≤  ≤ 1. This requires a nonexisting polyphase subfilter, denoted as p (k), which would “fall” between two existing subfilters, say, pi (k) and pi+1 (k). This set of coefficients would create a row between the rows with indexes i and i + 1. We note that each column of (8.4) consists of a segment of I consecutive samples of

800

Multirate Digital Signal Processing

the impulse response h(n) and covers one sampling interval Tx . Suppose next that we can approximate the coefficient set in each column by an L-degree polynomial Bk () =

L 

b(k)  ,

k = 0, 1, . . . , K − 1

(8.5)

=0

We note that evaluating (8.5) at  = i/I will provide the coefficients of p i(k)ψ polyphase subfilter. The type of the polynomial (Lagrange, Chebyschev, etc.) and the order L can be chosen to avoid any performance degradation compared to the original filter h(n). The sample at location t = nTx + Tx is determined by y((n + )Tx ) =

K−1 

Bk ()x((n − k)Tx ),

0≤≤1

(8.6)

k=0

where the required filter coefficients are computed using (8.5). If we substitute the polynomials (8.5) into the filtering formula (8.6) and we change the order of summations, we obtain y((n + )Tx ) =

K−1 L 

b(k)  x((n − k)Tx )

k=0 =0

=

L  =0



K−1 

b(k) x((n − k)Tx )

k=0

The last equation can be written as y((n + )Tx ) =

L 

v()

(8.7)

=0

where v() =

K−1 

b(k) x((n − k)Tx ),

 = 0, 1, . . . , L

(8.8)

k=0

Equation (8.7) can be interpreted as a Taylor series representation of the output sequence, where the terms v() are successive local derivatives determined from the input sequence. Relation (8.8) can be implemented using FIR filtering structures with system functions K−1  (k) (8.9) b z−k H (z) = k=0

The most efficient computation of polynomial (8.7) can be done using the nested Horner’s rule, which is illustrated next for L = 4: y() = c0 + c1  + c2 2 + c3 3 + c4 4 = c0 + (c1 + (c2 + (c3 + c4 )))

(8.10)

801

Multirate Digital Signal Processing

x(nTx)

HL(z)

...

HL−1(z) v(L−1)

v(L)

H1(z)

H0(z)

v(1)

v(0)

... y(nTx+∆Tx) ∆ Figure 8.1 Block diagram of the Farrow structure for sampling rate change by an arbitrary factor.

This approach leads to the block diagram realization shown in Figure 8.1, which is known as Farrow structure (Farrow 1988). Basically, the Farrow structure performs interpolation between signal values by interpolating between filter coefficients. More details can be found in Gardner (1993), Erup et al. (1993), Ramstad (1984), Harris (1997), and Laakso et al. (1996).

9

Applications of Multirate Signal Processing There are numerous practical applications of multirate signal processing. In this section we describe a few of these applications.

9.1

Design of Phase Shifters

Suppose that we wish to design a network that delays the signal x(n) by a fraction of a sample. Let us assume that the delay is a rational fraction of a sampling interval Tx [i.e., d = (k/I )Tx , where k and I are relatively prime positive integers]. In the frequency domain, the delay corresponds to a linear phase shift of the form (ω) = −

kω I

(9.1)

The design of an all-pass linear-phase filter is relatively difficult. However, we can use the methods of sample-rate conversion to achieve a delay of (k/I )Tx , exactly, without introducing any significant distortion in the signal. To be specific, let us consider the system shown in Fig. 9.1. The sampling rate is increased by a factor I using a standard interpolator. The lowpass filter eliminates the images in the spectrum of the interpolated signal, and its output is delayed by k samples at the sampling rate I Fx . The delayed signal is decimated by a factor D = I . Thus we have achieved the desired delay of (k/I )Tx . x(n) Fx

↑I

IFx

Lowpass filter

IFx

Delay by k samples

IFx

↓I

Figure 9.1 Method for generating a delay in a discrete-time signal.

802

y(n) Fx

Multirate Digital Signal Processing

x(n)

p0(n)

p1(n)

p2(n)



Output

Figure 9.2

Polyphase filter structure for implementing the system shown in Fig. 9.1.

Rate = IFx



pk(n)

pI−1(n)

An efficient implementation of the interpolator is the polyphase filter illustrated in Fig. 9.2. The delay of k samples is achieved by placing the initial position of the commutator at the output of the kth subfilter. Since decimation by D = I means that we take one out of every I samples from the polyphase filter, the commutator position can be fixed to the output of the kth subfilter. Thus a delay in k/I can be achieved by using only the kth subfilter of the polyphase filter. We note that the polyphase filter introduces an additional delay of (M − 1)/2 samples, where M is the length of its impulse response. Finally, we mention that if the desired delay is a nonrational factor of the sample interval Tx , the methods described in Section 8 can be used to obtain the delay.

9.2

Interfacing of Digital Systems with Different Sampling Rates

In practice we frequently encounter the problem of interfacing two digital systems that are controlled by independently operating clocks. An analog solution to this problem is to convert the signal from the first system to analog form and then resample it at the input to the second system using the clock in this system. However, a simpler approach is one where the interfacing is done by a digital method using the basic sample-rate conversion methods described in this chapter. To be specific, let us consider interfacing the two systems with independent clocks as shown in Fig. 9.3. The output of system A at rate Fx is fed to an interpolator which increases the sampling rate by I . The output of the interpolator is fed at the rate I Fx to a digital sample-and-hold which serves as the interface to system B at the high sampling rate I Fx . Signals from the digital sample-and-hold are read out into system B at the clock rate DFy of system B. Thus the output rate from the sample-and-hold is not synchronized with the input rate.

803

Multirate Digital Signal Processing

System A

x(n)

↑I Interpolation

IFx

Digital sampleand-hold

IFy

x(m)

↓D Decimator

Fx

DFy Clock A

IFx

System B

Fy Clock B

Figure 9.3 Interfacing of two digital systems with different sampling rates.

In the special case where D = I and the two clock rates are comparable but not identical, some samples at the output of the sample-and-hold may be repeated or dropped at times. The amount of signal distortion resulting from this method can be kept small if the interpolator/decimator factor is large. By using linear interpolation in place of the digital sample-and-hold we can further reduce the distortion and thus reduce the size of the interpolator factor.

9.3

Implementation of Narrowband Lowpass Filters

In Section 6 we demonstrated that a multistage implementation of sampling-rate conversion often provides for a more efficient realization, especially when the filter specifications are very tight (e.g., a narrow passband and a narrow transition band). Under similar conditions, a lowpass, linear-phase FIR filter may be more efficiently implemented in a multistage decimator-interpolator configuration. To be more specific, we can employ a multistage implementation of a decimator of size D , followed by a multistage implementation of an interpolator of size I , where I = D . We demonstrate the procedure by means of an example for the design of a lowpass filter which has the same specifications as the filter that is given in Example 6.1. EXAMPLE 9.1 Design a linear-phase FIR filter that satisfies the following specifications: Sampling frequency: Passband: Transition band: Stopband

8000 Hz 0 ≤ F ≤ 75 Hz 75 Hz ≤ F ≤ 80 Hz 80 Hz ≤ F ≤ 4000 Hz

Passband ripple:

δ1 = 10−2

Stopband ripple:

δ2 = 10−4

Solution. If this filter were designed as a single-rate linear-phase FIR filter, the length of the filter required to meet the specifications is (from Kaiser’s formula) ˆ ≈ 5152 M

804

Multirate Digital Signal Processing

Now, suppose that we employ a multirate implementation of the lowpass filter based on a decimation and interpolation factor of D = I = 100. A single-stage implementation of the decimator-interpolator requires an FIR filter of length ˆ 1 = −10 log10 (δ1 δ2 /2) − 13 + 1 ≈ 5480 M 14.6f However, there is a significant savings in computational complexity by implementing the decimator and interpolator filters using their corresponding polyphase filters. If we employ linear-phase (symmetric) decimation and interpolation filters, the use of polyphase filters reduces the multiplication rate by a factor of 100. A significantly more efficient implementation is obtained by using two stages of decimation followed by two stages of interpolation. For example, suppose that we select D1 = 50, D2 = 2, I1 = 2, and I2 = 50. Then the required filter lengths are ˆ 1 = −10 log(δ1 δ2 /4) − 13 + 1 ≈ 177 M 14.6f ˆ 1 = −10 log10 (δ1 δ2 /4) − 13 + 1 ≈ 233 M 14.6f Thus we obtain a reduction in the overall filter length of 2(5480)/2(177 + 233) ≈ 13.36. In addition, we obtain further reduction in the multiplication rate by using polyphase filters. For the first stage of decimation, the reduction in multiplication rate is 50, while for the second stage the reduction in multiplication rate is 100. Further reductions can be obtained by increasing the number of stages of decimation and interpolation.

9.4

Subband Coding of Speech Signals

A variety of techniques have been developed to efficiently represent speech signals in digital form for either transmission or storage. Since most of the speech energy is contained in the lower frequencies, we would like to encode the lower-frequency band with more bits than the high-frequency band. Subband coding is a method where the speech signal is subdivided into several frequency bands and each band is digitally encoded separately. An example of a frequency subdivision is shown in Fig. 9.4(a) . Let us assume that the speech signal is sampled at a rate Fs samples per second. The first frequency subdivision splits the signal spectrum into two equal-width segments, a lowpass signal (0 ≤ F ≤ Fs /4) and a highpass signal (Fs /4 ≤ F ≤ Fs /2). The second frequency subdivision splits the lowpass signal from the first stage into two equal bands, a lowpass signal (0 < F ≤ Fs /8) and a highpass signal (Fs /8 ≤ F ≤ Fs /4). Finally, the third frequency subdivision splits the lowpass signal from the second stage into two equal-bandwidth signals. Thus the signal is subdivided into four frequency bands, covering three octaves, as shown in Fig. 9.4(b).

805

Multirate Digital Signal Processing

Lowpass filter Lowpass filter

Highpass filter

Highpass filter

Decimator D=2

Encoder

Decimator D=2

Encoder

To channel

Highpass filter

Decimator D=2

Encoder

To channel

Decimator D=2

Decimator D=2

Speech signal

Lowpass filter

Decimator D=2

To channel

Encoder

To channel (a)

1 0

2 π 8

3 π 4

4 π 2

π

ω

(b)

Figure 9.4 Block diagram of a subband speech coder.

Decimation by a factor of 2 is performed after frequency subdivision. By allocating a different number of bits per sample to the signal in the four subbands, we can achieve a reduction in the bit rate of the digitalized speech signal. Filter design is particularly important in achieving good performance in subband coding. Aliasing resulting from decimation of the subband signals must be negligible. It is clear that we cannot use brickwall filter characteristics as shown in Fig. 9.5 (a), since such filters are physically unrealizable. A particularly practical solution to the aliasing problem is to use quadrature mirror filters (QMF), which have the frequency response characteristics shown in Fig. 9.5 (b). These filters are described in Section 11. The synthesis method for the subband encoded speech signal is basically the reverse of the encoding process. The signals in adjacent lowpass and highpass frequency bands are interpolated, filtered, and combined as shown in Fig. 9.6. A pair of QMF is used in the signal synthesis for each octave of the signal. Subband coding is also an effective method to achieve data compression in image signal processing. By combining subband coding with vector quantization for each subband signal, Safranek et al. (1988) have obtained coded images with approximately 21 bit per pixel, compared with 8 bits per pixel for the uncoded image. In general, subband coding of signals is an effective method for achieving bandwidth compression in a digital representation of the signal, when the signal energy is concentrated in a particular region of the frequency band. Multirate signal processing notions provide efficient implementations of the subband encoder.

806

Multirate Digital Signal Processing

Figure 9.5

Filter characteristics for subband coding.

Decoder

↑2

Filter

+

Decoder

↑2

Filter

Decoder

↑2

Filter

↑2

Filter

Decoder

↑2

Filter

+

↑2

Filter

+

Output

Figure 9.6 Synthesis of subband-encoded signals.

807

Multirate Digital Signal Processing

10

Digital Filter Banks Filter banks are generally categorized as two types, analysis filter banks and synthesis filter banks. An analysis filter bank consists of a set of filters, with system functions {Hk (z)}, arranged in a parallel bank as illustrated in Fig. 10.1(a). The frequency response characteristics of this filter bank split the signal into a corresponding number of subbands. On the other hand, a synthesis filter bank consists of a set of filters with system functions {Gk (z)}, arranged as shown in Fig. 10.1(b), with corresponding inputs {yk (n)}. The outputs of the filters are summed to form the synthesized signal {x(n)}. Filter banks are often used for performing spectrum analysis and signal synthesis. When a filter bank is employed in the computation of the discrete Fourier transform (DFT) of a sequence {x(n)}, the filter bank is called a DFT filter bank. An analysis filter bank consisting of N filters {Hk (z), k = 0, 1, . . . , N − 1} is called a uniform DFT filter bank if Hk (z), k = 1, 2, . . . , N − 1, are derived from a prototype filter

y1(n)

H1(z)

y2(n)



x(n)

H0(z)

yN−1(n)

HN−1(z) Analysis filter bank (a)

G0(z)

y2(n)

G1(z)

+



y1(n)

yN−1(n)

GN−1(z) Synthesis filter bank (b)

Figure 10.1 A digital filter bank.

808

x(n)

Multirate Digital Signal Processing

H0(ω)

H1(ω)

H2(ω)

HN−1(ω)

......

... −π N

0

π N

2π N

3π N

4π N

5π N

ω 2π(N−1) N



Figure 10.2 Illustration of frequency response characteristics of the N fil-

ters.

H0 (z), where   2π k , Hk (ω) = H0 ω − N

k = 1, 2, . . . , N − 1

(10.1)

Hence the frequency response characteristics of the filters {Hk (z), k = 0, 1, . . . , N −1} are simply obtained by uniformly shifting the frequency response of the prototype filter by multiples of 2π/N . In the time domain the filters are characterized by their impulse responses, which can be expressed as hk (n) = h0 (n)ej 2πnk/N ,

k = 0, 1, . . . , N − 1

(10.2)

where h0 (n) is the impulse response of the prototype filter, which in general may be either an FIR or an IIR filter. If H0 (z) denotes the transfer function of the prototype filter, the transfer function of the kth filter is Hk (z) = H0 (ze−j 2πk/N ),

1≤k ≤N −1

(10.3)

Figure 10.2 provides a conceptual illustration of the frequency response characteristics of the N filters. The uniform DFT analysis filter bank can be realized as shown in Fig. 10.3(a), where the frequency components in the sequence {x(n)} are translated in frequency to lowpass by multiplying x(n) with the complex exponentials exp(−j 2π nk/N ), k = 1, . . . , N −1, and the resulting product signals are passed through a lowpass filter with impulse response h0 (n). Since the output of the lowpass filter is relatively narrow in bandwidth, the signal can be decimated by a factor D ≤ N . The resulting decimated output signal can be expressed as Xk (m) =

 n

h0 (mD − n)x(n)e−j 2πnk/N ,

k = 0, 1, . . . , N − 1 m = 0, 1, . . .

(10.4)

where {Xk (m)} are samples of the DFT at frequencies ωk = 2π k/N .

809

Multirate Digital Signal Processing

e−jω0n h0(n)

↓D

X0(m)

h0(n)

↓D

X1(m)

h0(n)

↓D

h0(n)

↓D

x(n)





e−jω1n



e−jωk n



e−jωN−1n

Xk(m)

X N−1(m)

Analysis (a) e jω0n ↑D

g0(n) …

Y0(m)

g0(n)

↑D

g0(n)

+





YN−1(m)

↑D …

Y1(m)

e jω1n ν(n)

e jωN−1n

Synthesis (b)

ωk =

2πk N

Figure 10.3 A uniform DFT filter bank.

The corresponding synthesis filter for each element in the filter bank can be viewed as shown in Fig. 10.3 (b), where the input signal sequences {Yk (m), k = 0, 1, . . . , N − 1} are upsampled by a factor of I = D, filtered to remove the images, and translated in frequency by multiplication by the complex exponentials {exp(j 2πnk/N ), k = 0, 1, . . . , N − 1}. The resulting frequency-translated signals from the N filters are then summed. Thus we obtain the sequence   N−1 1  j 2πnk/N  e Yk (m)g0 (n − mI ) v(n) = N m k=0

=

 m

=

 m

810



N−1 1  g0 (n − mI ) Yk (m)ej 2πnk/N N k=0

g0 (n − mI )yn (m)

 (10.5)

Multirate Digital Signal Processing

where the factor 1/N is a normalization factor, {yn (m)} represent samples of the inverse DFT sequence corresponding to {Yk (m)}, {g0 (n)} is the impulse response of the interpolation filter, and I = D . The relationship between the output {Xk (n)} of the analysis filter bank and the input {Yk (m)} to the synthesis filter bank depends on the application. Usually, {Yk (m)} is a modified version of {Xk (m)}, where the specific modification is determined by the application. An alternative realization of the analysis and synthesis filter banks is illustrated in Fig. 10.4. The filters are realized as bandpass filters with impulse responses hk (n) = h0 (n)ej 2πnk/N ,

k = 0, 1, . . . , N − 1

(10.6)

The output of each bandpass filter is decimated by a factor D and multiplied by exp(−j 2πmk/N ) to produce the DFT sequence {Xk (m)}. The modulation by the complex exponential allows us to shift the spectrum of the signal from ωk = 2π k/N to ω0 = 0. Hence this realization is equivalent to the realization given in Fig. 10.3. e−jω0mD h0(n)

↓D

X0(m) e−jω1mD

hN−1(n)

X1(m) e−jωN−1mD

↓D





x(n)

↓D



h1(n)

XN−1(m)

Analysis (a) e jω0 mD Y0(m)

↑D

g0(n)

↑D

g1(n)

e jωN−1mD YN−1(m)



Y1(m)



e jω1mD

↑D

gN−1(n)

+

ν(n)

Synthesis (b)

Figure 10.4 Alternative realization of a uniform DFT filter bank.

811

Multirate Digital Signal Processing

The analysis filter bank output can be written as    x(n)h0 (mD − n)ej 2πk(mD−n)/N e−j 2πmkD/N Xk (m) =

(10.7)

n

The corresponding filter bank synthesizer can be realized as shown in Fig. 10.4(b), where the input sequences are first multiplied by the exponential factors [exp(j 2π kmD/N)], upsampled by the factor I = D, and the resulting sequences are filtered by the bandpass interpolation filters with impulse responses gk (n) = g0 (n)ej 2πnk/N

(10.8)

where {g0 (n)} is the impulse response of the prototype filter. The outputs of these filters are then summed to yield   N−1 1   j 2πkmI /N v(n) = [Yk (m)e ]gk (n − mI ) (10.9) N m k=0

where I = D. In the implementation of digital filter banks, computational efficiency can be achieved by use of polyphase filters for decimation and interpolation. Of particular interest is the case where the decimation factor D is selected to be equal to the number N of frequency bands. When D = N , we say that the filter bank is critically sampled.

10.1

Polyphase Structures of Uniform Filter Banks

For the analysis filter bank, let us define a set of N = D polyphase filters with impulse responses pk (n) = h0 (nN − k), k = 0, 1, . . . , N − 1 (10.10) and the corresponding set of decimated input sequences xk (n) = x(nN + k),

k = 0, 1, . . . , N − 1

(10.11)

Note that this definition of {pk (n)} implies that the commutator for the decimator rotates clockwise. The structure of the analysis filter bank based on the use of polyphase filters can be obtained by substituting (10.10) and (10.11) into (10.7) and rearranging the summation into the form   N−1   Xk (m) = (10.12) pn (l)xn (m − l) e−j 2πnk/N , k = 0, 1, . . . , D − 1 n=0

l

where N = D. Note that the inner summation represents the convolution of {pn (l)} with {xn (l)}. The outer summation represents the N -point DFT of the filter outputs. The filter structure corresponding to this computation is illustrated in Fig. 10.5. Each sweep of the commutator results in N outputs, denoted as {rn (m), n = 0, 1, . . . , N − 1} from the N polyphase filters. The N -point DFT of this sequence yields the spectral samples {Xk (m)}. For large values of N , the FFT algorithm provides an efficient means for computing the DFT.

812

Multirate Digital Signal Processing

x0(m)

p0(m)

r0(m)

X0(m)

p1(m)

r1(m)

X1(m)

m=0 x1(m)





pn(m)

xN−1(m)



N-Point DFT rn(m)

p N−1(m)

X k(m) …

xn(m)





x(l)

r N−1(m)

X N−1(m)

Figure 10.5 Digital filter bank structure for the computation of

(10.12).

Now suppose that the spectral samples {Xk (m)} are modified in some manner, prescribed by the application, to produce {Yk (m)}. A filter bank synthesis filter based on a polyphase filter structure can be realized in a similar manner. First, we define the impulse response of the N (D = I = N ) polyphase filters for the interpolation filter as qk (n) = g0 (nN + k), k = 0, 1, . . . , N − 1 (10.13) and the corresponding set of output signals as vk (n) = v(nN + k),

k = 0, 1, . . . , N − 1

(10.14)

Note that this definition of {qk (n)} implies that the commutator for the interpolator rotates counterclockwise. By substituting (10.13) into (10.5), we can express the output vl (n) of the l th polyphase filter as   N−1  1  j 2πkl/N (10.15) , l = 0, 1, . . . , N − 1 ql (n − m) Yk (m)e vl (n) = N m k=0

The term in brackets is the N -point inverse DFT of {Yk (m)}, which we denote as {yl (m)}. Hence  (10.16) vl (n) = ql (n − m)yl (m), l = 0, 1, . . . , N − 1 m

The synthesis structure corresponding to (10.16) is shown in Fig. 10.6. It is interesting to note that by defining the polyphase interpolation filter as in (10.13), the structure in Fig. 10.6 is the transpose of the polyphase analysis filter shown in Fig. 10.5.

813

Multirate Digital Signal Processing

Y0(m)

y0(m)

Y1(m)

y1(m)

q0(m)

ν0(m)

m=0 q1(m)

ν1(m)



yl (m)

YN−1(m)

yN−1(m)

… ql (m)

νl (m)



Inverse DFT



Yk(m)





ν(n)

qN−1(m)

νN−1(m)

Figure 10.6 Digital filter bank structure for the computa-

tion of (10.16).

In our treatment of digital filter banks we considered the important case of critically sampled DFT filter banks, where D = N . Other choices of D and N can be employed in practice, but the implementation of the filters becomes more complex. Of particular importance is the oversampled DFT filter bank, where N = KD , D denotes the decimation factor and K is an integer that specifies the oversampling factor. In this case it can be shown that the polyphase filter bank structures for the analysis and synthesis filters can be implemented by use of N subfilters and N -point DFTs and inverse DFTs.

10.2

Transmultiplexers

An application of digital filter banks is in the design and implementation of digital transmultiplexers, which are devices for converting between time-division-multiplexed (TDM) signals and frequency-division-multiplexed (FDM) signals. In a transmultiplexer for TDM-to-FDM conversion, the input signal {x(n)} is a time-division-multiplexed signal consisting of L signals, which are separated by a commutator switch. Each of these L signals is then modulated on a different carrier frequency to obtain an FDM signal for transmission. In a transmultiplexer for FDMto-TDM conversion, the composite signal is separated by filtering into the L signal components which are then time-division multiplexed. In telephony, single-sideband transmission is used with channels spaced at a nominal 4-kHz bandwidth. Twelve channels are usually stacked in frequency to form a basic group channel, with a bandwidth of 48 kHz. Larger-bandwidth FDM signals are formed by frequency translation of multiple groups into adjacent frequency bands. We shall confine our discussion to digital transmultiplexers for 12-channel FDM and TDM signals. Let us first consider FDM-to-TDM conversion. The analog FDM signal is passed through an A/D converter as shown in Fig. 10.7(a). The digital signal is then

814

Multirate Digital Signal Processing

SSB demodulator

Decimator

s1(n)

s2(n) TDM Multiplexer



A/D converter

Decimator

TDM signal



FDM signal

SSB demodulator

SSB demodulator

Decimator

sN (n)

(a) cos ωkn

x(n)

LPF h(n)

↓D

LPF h(n)

↓D

−sin ωkn

(b)

Figure 10.7 Block diagram of FDM-to-TDM transmultiplexer.

demodulated to baseband by means of single-sideband demodulators. The output of each demodulator is decimated and fed to the commutator of the TDM system. To be specific, let us assume that the 12-channel FDM signal is sampled at the Nyquist rate of 96 kHz and passed through a filter-bank demodulator. The basic building block in the FDM demodulator consists of a frequency converter, a lowpass filter, and a decimator, as illustrated in Fig. 10.7(b). Frequency conversion can be efficiently implemented by the DFT filter bank described previously. The lowpass filter and decimator are efficiently implemented by use of the polyphase filter structure. Thus the basic structure for the FDM-to-TDM converter has the form of a DFT filter bank analyzer. Since the signal in each channel occupies a 4-kHz bandwidth, its Nyquist rate is 8 kHz, and hence the polyphase filter output can be decimated by a factor of 12. Consequently, the TDM commutator is operating at a rate of 12 ×8 kHz or 96 kHz. In TDM-to-FDM conversion, the 12-channel TDM signal is demultiplexed into the 12 individual signals, where each signal has a rate of 8 kHz. The signal in each channel is interpolated by a factor of 12 and frequency converted by a single-sideband modulator, as shown in Fig. 10.8. The signal outputs from the 12 single-sideband modulators are summed and fed to the D/A converter. Thus we obtain the analog

815

Multirate Digital Signal Processing

Interpolator

SSB modulator

TDM signal

+

… …

SSB modulator

… …

Interpolator

Interpolator

FDM D/A convertor signal

SSB modulator

Figure 10.8 Block diagram of TDM-to-FDM transmultiplexer.

FDM signal for transmission. As in the case of FDM-to-TDM conversion, the interpolator and the modulator filter are combined and efficiently implemented by use of a polyphase filter. The frequency translation can be accomplished by the DFT. Consequently, the TDM-to-FDM converter encompasses the basic principles introduced previously in our discussion of DFT filter bank synthesis.

11

Two-Channel Quadrature Mirror Filter Bank The basic building block in applications of quadrature mirror filters (QMF) is the twochannel QMF bank shown in Fig. 11.1. This is a multirate digital filter structure that employs two decimators in the “signal analysis” section and two interpolators in the “signal synthesis” section. The lowpass and highpass filters in the analysis section have impulse responses h0 (n) and h1 (n), respectively. Similarly, the lowpass and highpass filters contained in the synthesis section have impulse responses g0 (n) and g1 (n), respectively The Fourier transforms of the signals at the outputs of the two decimators are  

   ω

ω − 2π 1 ω ω − 2π X H0 +X H0 Xa0 (ω) = 2 2 2 2 2 (11.1)  

  

ω ω − 2π 1 ω ω − 2π Xa1 (ω) = X H1 +X H1 2 2 2 2 2 ↓2

Xa0

~ ~

LPF H0(z)

Xs0

↑2

G0(z)

x(n)

+

↓2

Xa1

~ ~

HPF H1(z) Analysis section

Figure 11.1 Two-channel QMF bank.

816

Xs1

↑2

G1(z) Synthesis section

ˆ x(n)

Multirate Digital Signal Processing

If Xs0 (ω) and Xs1 (ω) represent the two inputs to the synthesis section, the output is simply ˆ X(ω) = Xs0 (2ω)G0 (ω) + Xs1 (2ω)G1 (ω) (11.2) Now, suppose that we connect the analysis filter to the corresponding synthesis filter, so that Xa0 (ω) = Xs0 (ω) and Xa1 (ω) = Xs1 (ω). Then, by substituting from (11.1) into (11.2), we obtain 1 ˆ X(ω) = [H0 (ω)G0 (ω) + H1 (ω)G1 (ω)] X(ω) 2 1 + [H0 (ω − π )G0 (ω) + H1 (ω − π )G1 (ω)] X(ω − π ) 2

(11.3)

The first term in (11.3) is the desired signal output from the QMF bank. The second term represents the effect of aliasing, which we would like to eliminate. In the z-transform domain (11.3) is expressed as 1 ˆ X(z) = [H0 (z)G0 (z) + H1 (z)G1 (z)] X(z) 2 +

1 [H0 (−z)G0 (z) + H1 (−z)G1 (z)] X(−z) 2

(11.4)

= Q(z)X(z) + A(z)X(−z) where, by definition, Q(z) =

1 [H0 (z)G0 (z) + H1 (z)G1 (z)] 2

1 A(z) = [H0 (−z)G0 (z) + H1 (−z)G1 (z)] 2

11.1

(11.5)

Elimination of Aliasing

To eliminate aliasing, we require that A(z) = 0, i.e., H0 (−z)G0 (z) + H1 (−z)G1 (z) = 0

(11.6)

In the frequency domain, this condition becomes H0 (ω − π )G0 (ω) + H1 (ω − π )G1 (ω) = 0

(11.7)

This condition can be simply satisfied by selecting G0 (ω) and G1 (ω) as G0 (ω) = H1 (ω − π ),

G1 (ω) = −H0 (ω − π )

(11.8)

Thus, the second term in (11.3) vanishes and the filter bank is alias-free.

817

Multirate Digital Signal Processing

H0(ω)

1

0

H1(ω)

π 2

π

ω

Figure 11.2 Mirror image characteristics of the analysis filters

H0 (ω) and H1 (ω).

To elaborate, let us assume that H0 (ω) is a lowpass filter and H1 (ω) is a mirrorimage highpass filter as shown in Fig. 11.2. Then we can express H0 (ω) and H1 (ω) as H0 (ω) = H (ω) (11.9) H1 (ω) = H (ω − π ) where H (ω) is the frequency response of a lowpass filter. In the time domain, the corresponding relations are h0 (n) = h(n) h1 (n) = (−1)n h(n)

(11.10)

As a consequence, H0 (ω) and H1 (ω) have mirror-image symmetry about the frequency ω = π/2, as shown in Fig.11.2. To be consistent with the constraint in (11.8), we select the lowpass filter G0 (ω) as G0 (ω) = H (ω)

(11.11)

G1 (ω) = −H (ω − π )

(11.12)

and the highpass filter G1 (ω) as

In the time domain, these relations become g0 (n) = h(n) g1 (n) = (−1)n h(n)

(11.13)

In the z-transform domain, the relations for the elimination of aliasing are: H0 (z) = H (z) H1 (z) = H (−z) G0 (z) = H (z) G1 (z) = −H (−z)

818

(11.14)

Multirate Digital Signal Processing

11.2

Condition for Perfect Reconstruction

With A(z) = 0, we now consider the condition for which the output x(n) of the QMF bank is identical to the input x(n), except for an arbitrary delay, for all possible inputs. When this condition is satisfied, the filter bank is called a perfect reconstruction QMF bank. Thus, we require that Q(z) =

1 [H0 (z)G0 (z) + H1 (z)G1 (z)] = z−k 2

(11.15)

By making use of the relations in (11.14), the condition for perfect reconstruction may be expressed as H 2 (z) − H 2 (−z) = 2z−k (11.16) or, equivalently,

H 2 (ω) − H 2 (ω − π ) = 2e−j ωk

(11.17)

Therefore, for perfect reconstruction, the frequency response H (ω) of the lowpass filter in the two-channel QMF bank must satisfy the magnitude condition     2 H (ω) − H 2 (ω − π ) = C

(11.18)

where C is a positive constant, e.g., C = 2. We note that if H (ω) satisfies the magnitude condition in (11.18) and is designed to have linear phase, then the QMF output x(n) is simply a delayed version of the input sequence x(n). However, linear phase is not a necessary condition for perfect reconstruction.

11.3

Polyphase Form of the QMF Bank

The two-channel alias-free QMF bank can be realized efficiently by employing polyphase filters. Toward this goal, H0 (z) and H1 (z) can be expressed as H0 (z) = P0 (z2 ) + z−1 P1 (z2 ) H1 (z) = P0 (z2 ) − z−1 P1 (z2 )

(11.19)

where we have used the relationship given in (11.14). Similarly, using the relations given in (11.14), we obtain the polyphase representation of the filters G0 (z) and G1 (z) as G0 (z) = P0 (z2 ) + z−1 P1 (z2 )   G1 (z) = − P0 (z2 ) − z−1 P1 (z2 )

(11.20)

Thus, we obtain the polyphase realization of the QMF bank as shown in Figure 11.3(a). The corresponding computationally efficient polyphase realization is shown in Figure 11.3(b).

819

Multirate Digital Signal Processing

x(n)

+

P0(z2)

↑2

↓2

+

P1(z2) z−1

z−1 P1(z2)



+

ˆ x(n)

↑2

↓2

+

P0(z2)

P1(z2)

↑2



+

(a) x(n)

↓2

P0(z2)

+

+

z−1

z−1 ↓2

P1(z2)



+



Analysis Section

+

P0(z2)

↑2

ˆ x(n) +

Synthesis Section (b)

Figure 11.3 Polyphase realization of the two-channel QMF bank.

11.4

Linear Phase FIR QMF Bank

Now, let us consider the use of a linear phase filter H (ω). Hence H (ω) may be expressed in the form (11.21) H (ω) = Hr (ω)e−j ω(N −1)/2 where N is the filter length. Then H 2 (ω) = Hr2 (ω)e−j ω(N −1) = |H (ω)|2 e−j ω(N −1) and

H 2 (ω − π) = Hr2 (ω − π )e−j (ω−π)(N−1) = (−1)N−1 |H (ω − π )|2 e−j ω(N −1)

(11.22)

(11.23)

Therefore, the overall transfer function of the two-channel QMF which employs linear-phase FIR filters is   ˆ X(ω) = |H (ω)|2 − (−1)N−1 |H (ω − π )|2 e−j ω(N −1) X(ω)

(11.24)

Note that the overall filter has a delay of N −1 samples and a magnitude characteristic M(ω) = |H (ω)|2 − (−1)N−1 |H (ω − π )|2

(11.25)

We also note that when N is odd, M(π/2) = 0, because |H (π/2)| = |H (3π/2)|. This is an undesirable property for a QMF design. On the other hand, when N is even, M(ω) = |H (ω)|2 + |H (ω − π )|2

820

(11.26)

Multirate Digital Signal Processing

which avoids the problem of a zero at ω = π/2. For N even, the ideal two-channel QMF should satisfy the condition M(ω) = |H (ω)|2 + |H (ω − π )|2 = 1

for all ω

(11.27)

which follows from (11.25). Unfortunately, the only filter frequency response function that satisfies (11.27) is the trivial function |H (ω)| 2 = cos2 aω. Consequently, any nontrivial linear-phase FIR filter H (ω) introduces some amplitude distortion. The amount of amplitude distortion introduced by a nontrivial linear phase FIR filter in the QMF can be minimized by optimizing the FIR filter coefficients. A particularly effective method is to select the filter coefficients of H (ω) such that M(ω) is made as flat as possible while simultaneously minimizing (or constraining) the stopband energy of H (ω). This approach leads to the minimization of the integral squared error J =w

π



π

|H (ω)|2 dω + (1 − w)

[M(ω) − 1]2 dω

(11.28)

0

ωs

where w is a weighting factor in the range 0 < w < 1. In performing the optimization, the filter impulse response is constrained to be symmetric (linear phase). This optimization is easily done numerically on a digital computer. This approach has been used by Johnston (1980) and Jain and Crochiere (1984) to design two-channel QMFs. Tables of optimum filter coefficients have been tabulated by Johnston (1980).

11.5

IIR QMF Bank

As an alternative to linear-phase FIR filters, we can design an IIR filter that satisfies the all-pass constraint given by (11.18). For this purpose, elliptic filters provide especially efficient designs. Since the QMF would introduce some phase distortion, the signal at the output of the QMF can be passed through an all-pass phase equalizer designed to minimize phase distortion.

11.6

Perfect Reconstruction Two-Channel FIR QMF Bank

We have observed that neither linear-phase FIR filters nor IIR filters described above yield perfect reconstruction in a two-channel QMF bank. However, as shown by Smith and Barnwell (1984), perfect reconstruction can be achieved by designing H (ω) as a linear-phase FIR half-band filter of length 2N − 1. A half-band filter is defined as a zero-phase FIR filter whose impulse response {b(n)} satisfies the condition  b(2n) =

constant, n = 0 0, n = 0

(11.29)

821

Multirate Digital Signal Processing

1 + δ1 1 − δ1

δ1 = δ2

ωp + ωs = π

δ2 0

ωp ωs −δ2

Figure 11.4 Frequency response characteristic of FIR half-band

filter.

Hence all the even-numbered samples are zero except at n = 0. The zero-phase requirement implies that b(n) = b(−n). The frequency response of such a filter is B(ω) =

K 

b(n)e−j ωn

(11.30)

n=−K

where K is odd. Furthermore, B(ω) satisfies the condition that B(ω) + B(π − ω) is equal to a constant for all frequencies. The typical frequency response characteristic of a half-band filter is shown in Fig. 11.4. We note that the filter response is symmetric with respect to π/2, the band edges frequencies ωp and ωs are symmetric about ω = π/2, and the peak passband and stopband errors are equal. We also note that the filter can be made causal by introducing a delay of K samples. Now, suppose that we design an FIR half-band filter of length 2N −1, where N is even, with frequency response as shown in Fig. 11.5 (a). From B(ω) we construct another half-band filter with frequency response B+ (ω) = B(ω) + δe−j ω(N −1)

(11.31)

as shown in Fig. 11.5(b). Note that B+ (ω) is nonnegative and hence it has the spectral factorization (11.32) B+ (z) = H (z)H (z−1 )z−(N −1) or, equivalently, B+ (ω) = |H (ω)|2 e−j ω(N −1)

(11.33)

where H (ω) is the frequency response of an FIR filter of length N with real coefficients. Due to the symmetry of B+ (ω) with respect to ω = π/2, we also have

822

Multirate Digital Signal Processing

1+δ

1

1−δ

B(ω) amplitude response of G(z)

ωp + ωs = π

δ 0

ωp

π 2

ωs

−δ

(a)

1 + 2δ 1 B+(ω) amplitude response of B+(z) 2δ 0

ωp

π 2

ω

ωs

(b)

δ Magnitude responses of analysis filters

H0

H1

ω

0



(c)

Figure 11.5 Frequency response characteristic of FIR half-band filters

B(ω) and B+ (ω). (From Vaidyanathan (1987))

B+ (z) + (−1)N−1 B+ (−z) = αz−(N −1)

(11.34)

B+ (ω) + (−1)N−1 B+ (ω − π ) = αe−j ω(N −1)

(11.35)

or, equivalently,

where α is a constant. Thus, by substituting (11.32) into (11.34), we obtain H (z)H (z−1 ) + H (−z)H (−z−1 ) = α

(11.36)

823

Multirate Digital Signal Processing

Since H (z) satisfies (11.36) and since aliasing is eliminated when we have G 0(z) = H1 (−z) and G1 (z) = −H0 (−z), it follows that these conditions are satisfied by choosing H1 (z), G0 (z), and G1 (z) as H0 (z) = H (z) H1 (z) = −z−(N −1) H0 (−z−1 ) G0 (z) = z−(N −1) H0 (z−1 ) G1 (z) = z−(N −1) H1 (z−1 ) = −H0 (−z)

(11.37)

ˆ Thus aliasing distortion is eliminated and, since X(ω)/X(ω) is a constant, the QMF performs perfect reconstruction so that x(n) = αx(n − N + 1). However, we note that H (z) is not a linear-phase filter. The FIR filters H0 (z), H1 (z), G0 (z), and G1 (z) in the two-channel QMF bank are efficiently realized as polyphase filters as shown previously.

11.7

Two-Channel QMF Banks in Subband Coding

In Section 9.4 we described a method for efficient encoding of a speech signal based on subdividing the signal into several subbands and encoding each subband separately. For example, in Figure 9.4 we illustrated the separation of a signal into four subbands, namely, 0 ≤ F ≤ Fs /16, Fs /16 < F ≤ Fs /8, Fs /8 < F ≤ Fs /4, and Fs /4 < F ≤ Fs /2, where Fs is the sampling frequency. The subdivision into four subbands can be accomplished by use of three two-channel QMF analysis sections. After encoding and transmission through the channel, each subband signal is decoded and reconstructed by passing the subbands through three two-channel QMF synthesis filters. The system configuration for subband coding employing four subbands is illustrated in Figure 11.6. Encoder

Decoder

Analysis QMF

Synthesis QMF Encoder

Analysis QMF

Encoder

Signal Analysis samples QMF

Decoder Synthesis QMF

Channel

Synthesis Output QMF signal ˆ x(n)

Decoder

x(n)

Encoder

Decoder (a)

x(n)

↓2

+

P0(z)

+

P1(z)

↑2 z−1

z−1 ↓2

P1(z)

Analysis Section of QMF (b)



+

↑2 + P0(z) − Synthesis Section of QMF (c)

Figure 11.6 System for subband coding using two-channel QMF banks.

824

+

ˆ x(n)

Multirate Digital Signal Processing

M -Channel QMF Bank In this section, we consider the generalization of the QMF bank to M channels. Figure 12.1 illustrates the structure of an M -channel QMF bank, where x(n) is the input to the analysis section, xk(a) (n), 0 ≤ k ≤ M − 1, are the outputs of the analysis filters, xk(s) (n), 0 ≤ k ≤ M − 1, are the inputs to the synthesis filters and x(n) ˆ is the output of the synthesis section. The M outputs from the analysis filters may be expressed in the z-transform domain as Xk(a) (z)

M−1



1  m m = Hk z1/M WM X z1/M WM , M

0≤k ≤M −1

(12.1)

m=0

where WM = e−j 2π/M . The output from the synthesis section is ˆ X(z) =

M−1 

  Xk(s) zM Gk (z)

(12.2)

k=0

As in the case of the two-channel QMF bank, we set Xk(a) (z) = Xk(s) (z). Then, if we substitute (12.1) into (12.2), we obtain ˆ X(z) =

M−1  k=0



 M−1  m  m 1  Gk (z) Hk zWM X zWM M m=0

(12.3)

  M−1  1 M−1   m  m = Gk (z)Hk zWM X zWM M

G1(z)

G2(z)

GM−1(z)

M

M

M

+

ˆ x(n) Xz

...

... ↓M

x(s) 2 (n)

(s) (a) xM−1 (n) xM−1(n)

~ ~

HM−1(z)

x(a)2 (n)

x(s) 1 (n)



↓M

x(a)1 (n)

G0(z)



H2(z)

x(a)0 (n)



↓M

~ ~

H1(z)

x(s) 0 (n)

~ ~

↓M

k=0

~ ~

x(n) x(z)

H0(z)

...

m=0



12

M

Figure 12.1 An M -channel QMF bank.

825

Multirate Digital Signal Processing

It is convenient to define the term in the brackets as Rm (z) =

M−1 1  m Gk (z)Hk (zWM ), M

0≤m≤M −1

(12.4)

k=0

Then, (12.3) may be expressed as ˆ X(z) =

M−1 

Rm (z)X(zWNm )

m=0

= R0 (z)X(z) +

M−1 

(12.5)  m Rm (z)X zWM

m=1

We note the first term in (12.5) is the alias-free component of the QMF bank and the second term is the aliasing component.

12.1

Alias-Free and Perfect Reconstruction Condition

From (12.5), it is clear that aliasing is eliminated by forcing the condition Rm (z) = 0,

1≤m≤M −1

(12.6)

With the elimination of the alias terms, the M -channel QMF bank becomes a linear time-invariant system that satisfies the input-output relation ˆ X(z) = R0 (z)X(z) where R0 (z) =

M−1 1  Hk (z)Gk (z) M

(12.7)

(12.8)

k=0

Then, the condition for a perfect reconstruction M -channel QMF bank becomes R0 (z) = Cz−k

(12.9)

where C and k are positive constants.

12.2

Polyphase Form of the M -Channel QMF Bank

An efficient implementation of the M -channel QMF bank is achieved by employing polyphase filters. To obtain the polyphase form for the analysis filter bank, the kth filter Hk (z) is represented as Hk (z) =

M−1  m=0

826

z−m Pkm (z),

0≤k ≤M −1

(12.10)

Multirate Digital Signal Processing

We may express the equations for the M polyphase filter in matrix form as

where

and



H(z) = P(zM )a(z)

(12.11)

H(z) = [H0 (z) H1 (z) · · · HM−1 (z)]t  t a(z) = 1 z−1 z−2 · · · z−(M−1)

(12.12)

P00 (z) P10 (z) .. .

P01 (z) P11 (z)

PM−1 0 (z)

PM−1 1 (z)

 P(z) =  

 P0M−1 (z) P1M−1 (z)   

··· ···

(12.13)

· · · PM−1 M−1 (z)

The polyphase form of the analysis filter bank is shown in Figure 12.2(a), and after applying the first noble identity we obtain the structure shown in Figure 12.2(b). The synthesis section can be constructed in a similar manner. Suppose we use a type II (transpose) form (see Problem 15) for the polyphase representation of the filters {Gk (z)}. Thus, Gk (z) =

M−1 

z−(M−1−m) Qkm (zM ),

0≤k ≤M −1

(12.14)

m=0

When expressed in matrix form, (12.14) becomes G(z) = z−(M−1) Q(zM )a(z−1 )

(12.15)

where a(z) is defined in (12.12) and G(z) = [ G0 (z) G1 (z) · · · GM−1 (z) ]t  Q00 (z) Q01 (z) ··· Q0 M−1 (z) Q11 (z) ··· Q1 M−1 (z)  Q10 (z) Q(z) =  .  ..

z−1 P(zM)

x(a) 0 (n)

↓M

x 1 (n)

↓M

(a)

x(a) 2 (n)

...

...

z−1

↓M

z−1

↓M

(a)

x(n) z−1

(12.16)

· · · QM−1 M−1 (z) ↓M

x(a) 0 (n)

↓M

x 1 (n)

(a)

z−1

z−1 (a) xM−1 (n)

  

↓M

P(z)

x(a) 2 (n)

...

x(n)

QM−1 1 (z)

...

QM−1 0 (z)



↓M

(a) xM−1 (n)

(b)

Figure 12.2 Polyphase structure of the analysis section of an M -channel

QMF bank (a) before and (b) after applying the first noble identity.

827

Multirate Digital Signal Processing

x(s) 0 (n)



M



x(s) 0 (n)

M

z−1 x(s) 1 (n)



+

M



x(s) 1 (n)

z−1 +

M

z−1 Q(z)

+

...

...

M

z−1

z−1



+

M

x(s) 2 (n)

+

ˆ x(n)

(s) xM−1 (n)





(s) xM−1 (n)

M

z−1 ↓

x(s) 2 (n)

Q(zM)

+

M

(a)

ˆ x(n)

(b)

Figure 12.3 Polyphase structure of the synthesis section of an M -channel QMF

bank (a) before and (b) after applying the first noble identity.

Therefore, the synthesis section of the M-channel QMF bank is realized as shown in Figure 11.3. By combining Figures 12.2(b) and 12.3(b), we obtain the polyphase structure of the complete M-channel QMF bank shown in Figure 12.4. From the structure of the M -channel QMF bank shown in Figure 12.4, we observe that the perfect reconstruction condition can be restated as Q(z)P(z) = Cz−k I

(12.17)

where I is the M × M identity matrix. Hence, if the polyphase matrix P(z) is known, then the polyphase synthesis matrix Q(z) is Q(z) = Cz−k [P(z)]−1

(12.18)

↓M



M

z−1

z−1 ↓M



M

z−1 ↓M

P(z)

Q(z)



... ...

z−1

+

M

+ ...

x(n)

Figure 12.4

828

z−1

z−1 ↓M



Polyphase realization of the M -channel QMF bank

M

+

ˆ x(n)

Multirate Digital Signal Processing

EXAMPLE 12.1 Suppose the polyphase matrix for a three-channel perfect reconstruction FIR QMF bank is  1 1 2 P(z ) =  2 3 1  1 2 1 

3

Determine the analysis and the synthesis filters in the QMF bank. Solution.

The analysis filters are given by (12.11) as 

  H0 (z) 1  H1 (z)  =  2 1 H2 (z)

  1 1 2 −1 3 1z  2 1 z−2

Hence, H0 (z) = 1 + z−1 + 2z−2 ,

H1 (z) = 2 + 3z−1 + z−2 ,

The inverse of P(z3 ) is 

P(z3 )

−1

 1 1 = −1 2 1

H2 (z) = 1 + 2z−1 + z−2

 3 −5 −1 3 −1 1

We may scale this inverse by the factor 2, so that

Q(z3 ) = 2[P(z3 )]−1 Then, by applying (12.15), we obtain the synthesis filters as     1 G0 (z) 1 3 −5  G1 (z)  = z−2  −1 −1 3z  z2 1 −1 1 G2 (z) 

Hence, G0 (z) = −5 + 3z−1 + z−2 ,

G1 (z) = 3 − z−1 − z−2 ,

G3 (z) = 1 − z−1 + z−2

Vaidyanathan (1992) treats the design of M -channel perfect reconstruction QMF banks by selecting the analysis filters Hk (z) to be FIR with a paraunitary polyphase structure, i.e., ˜ P(z)P(z) = dI, d>0 (12.19) ˜ ˜ where P(z) is the paraconjugate of P(z). That is, P(z) is formed by the transpose of P(1/z), with the coefficients of P(z) replaced by their complex conjugates. Then, the polyphase filters in the synthesis section are designed to satisfy ˜ Q(z) = Cz−k P(z),

C > 0, k > 0

(12.20)

829

Multirate Digital Signal Processing

H0

1

H1

H2

Figure 12.5

Magnitude response for analysis filters in an M = 3 QMF bank.

π 3

0

ω

π

2π 3

It can be shown (see Vaidyanathan et al. (1989)) that any causal Lth-degree FIR paraunitary matrix P(z) can be expressed in a product form as P(z) = VL (z)VL−1 (z) · · · V1 (z)U

(12.21)

where U is a unitary matrix and {Vm (z)} are paraunitary matrices that have the form H H Vm (z) = I − υm υm + z−1 υm υm

(12.22)

where υm is an M -dimensional vector of unit norm. Then, the design of the synthesis filters reduces to the optimization of the components of υm and ui by minimizing an objective function, which may be a generalization of the two-channel QMF objective function given by (11.28). In particular, we may minimize the objective function J =

M−1  k=0

|Hk (ω)|2 dω

(12.23)

kth stopband

by employing a nonlinear optimization technique to determine the components of υm and ui . Thus, the vectors υm and ui completely determine P(z) and, hence, the analysis filters Hk (z). The synthesis filters are then determined from (12.20). For example, consider the design of a perfect reconstruction three-channel QMF bank with magnitude responses as shown in Fig. 12.5. The design procedure described above yields the magnitude responses for the analysis filters that are shown in Fig. 12.6. The filter length is N = 14. We observe that the stopband attenuation of the filters is approximately 20 dB. 0 H0(ω)

−10

Magnitude response of optimized analysis filters for an M = 3 FIR perfect reconstruction QMF bank. (From Multirate Systems and Filter Banks, by P.P. Vaidyanathan, ©1993 by Prentice Hall. Reprinted with permission of the publisher.)

830

H2(ω)

−20 −30

dB

Figure 12.6

H1(ω)

−40 −50 −60 0.0

0.1

0.2 0.3 Normalized frequency (ω/2π)

0.4

0.5

Multirate Digital Signal Processing

13

Summary and References The need for sampling rate conversion arises frequently in digital signal processing applications. In this chapter we first treated sampling rate reduction (decimation) and sampling rate increase (interpolation) by integer factors and then demonstrated how the two processes can be combined to obtain sampling rate conversion by any rational factor. Then, we described a method to achieve sampling rate conversion by an arbitrary factor. In the special case where the signal to be resampled is a bandpass signal, we described methods for performing the sampling rate conversion. In general, the implementation of sampling rate conversion requires the use of a linear time-variant filter. We described methods for implementing such filters, including the class of polyphase filter structures, which are especially simple to implement. We also described the use of multistage implementations of multirate conversion as a means of simplifying the complexity of the filter required to meet the specifications. We also described a number of applications that employ multirate signal processing, including the implementation of narrowband filters, phase shifters, filter banks, subband speech coders, quadrature mirror filters and transmultiplexers. These are just a few of the many applications encountered in practice where multirate signal processing is used. The first comprehensive treatment of multirate signal processing was given in the book by Crochiere and Rabiner (1983). In the technical literature, we cite the papers by Schafer and Rabiner (1973), and Crochiere and Rabiner (1975, 1976, 1981, 1983). The use of interpolation methods to achieve sampling rate conversion by an arbitrary factor is treated in a paper by Ramstad (1984). A thorough tutorial treatment of multirate digital filters and filter banks, including quadrature mirror filters, is given by Vetterli (1987), and by Vaidyanathan (1990, 1993), where many references on various applications are cited. A comprehensive survey of digital transmultiplexing methods is found in the paper by Scheuermann and Gockler (1981). Subband coding of speech has been considered in many publications. The pioneering work on this topic was done by Crochiere (1977, 1981) and by Garland and Esteban (1980). Subband coding has also been applied to coding of images. We mention the papers by Vetterli (1984), Woods and O’Neil (1986), Smith and Eddins (1988), and Safranek et al. (1988) as just a few examples. In closing, we wish to emphasize that multirate signal processing continues to be a very active research area.

Problems 1

An analog signal xa (t) is bandlimited to the range 900 ≤ F ≤ 1100 Hz. It is used as an input to the system shown in Fig. P1. In this system, H (ω) is an ideal lowpass filter with cutoff frequency Fc = 125 Hz. xa(t)

A/D converter Fx =

x(n)

w(n)

cos(0.8πn) 1 = 2500 Tx

H(ω)

v(n)

↓10

Fy =

y(n) 1 = 250 Ty

Figure P1

831

Multirate Digital Signal Processing

(a) Determine and sketch the spectra for the signals x(n), w(n), v(n), and y(n). (b) Show that it is possible to obtain y(n) by sampling xa (t) with period T = 4 milliseconds. 2 Consider the signal x(n) = a n u(n), |a| < 1. (a) Determine the spectrum X(ω). (b) The signal x(n) is applied to a decimator that reduces the rate by a factor of 2. Determine the output spectrum. (c) Show that the spectrum in part (b) is simply the Fourier transform of x(2n). 3 The sequence x(n) is obtained by sampling an analog signal with period T . From this signal a new signal is derived having the sampling period T /2 by use of a linear interpolation method described by the equation  n even  x(n/2),      n−1 n+1 y(n) = 1 x +x , n odd  2 2 2 (a) Show that this linear interpolation scheme can be realized by basic digital signal processing elements. (b) Determine the spectrum of y(n) when the spectrum of x(n) is  1, 0 ≤ |ω| ≤ 0.2π X(ω) = 0, otherwise (c) Determine the spectrum of y(n) when the spectrum of x(n) is  1, 0.7π ≤ |ω| ≤ 0.9π X(ω) = 0, otherwise 4 Consider a signal x(n) with Fourier transform X(ω) = 0,

ωn < |ω| ≤ π fm < |f | ≤ 21

(a) Show that the signal x(n) can be recovered from its samples x(mD) if the sampling frequency ωs = 2π/D ≤ 2ωm (fs = 1/D ≥ 2fm ). (b) Show that x(n) can be reconstructed using the formula x(n) =

∞ 

x(kD)hr (n − kD)

k=−∞

where

sin(2πfc n) fm < fc < fs − fm , ωm < ωc < ωs − ωm 2π n (c) Show that the bandlimited interpolation in part (b) can be thought of as a twostep process, first, increasing the sampling rate by a factor of D by inserting (D−1) zero samples between successive samples of the decimated signal xa (n) = x(nD), and second, filtering the resulting signal using an ideal lowpass filter with cutoff frequency ωc . hr (n) =

832

Multirate Digital Signal Processing

5

In this problem we illustrate the concepts of sampling and decimation for discretetime signals. To this end consider a signal x(n) with Fourier transform X(ω) as in Fig. P5. x(n)

X(ω) 1

… −2 −1 0 1 2 …

n

−π

−π 3

0

π 3

π ω

Figure P5

(a) Sampling x(n) with a sampling period D = 2 results in the signal  xs (n) =

x(n), 0,

n = 0, ±2, ±4, . . . n = ±1, ±3, ±5, . . .

Compute and sketch the signal xs (n) and its Fourier transform Xs (ω). Can we reconstruct x(n) from xs (n)? How? (b) Decimating x(n) by a factor of D = 2 produces the signal xd (n) = x(2n),

all n

Show that Xd (ω) = Xs (ω/2). Plot the signal xd (n) and its transform Xd (ω). Do we lose any information when we decimate the sampled signal xs (n)? 6

Design a decimator that downsamples an input signal x(n) by a factor D = 5. Use the Remez algorithm to determine the coefficients of the FIR filter that has a 0.1-dB ripple in the passband (0 ≤ ω ≤ π/5) and is down by at least 30 dB in the stopband. Also determine the corresponding polyphase filter structure for implementing the decimator.

7

Design an interpolator that increases the input sampling rate by a factor of I = 2. Use the Remez algorithm to determine the coefficients of the FIR filter that has a 0.1-dB ripple in the passband (0 ≤ ω ≤ π/2) and is down by at least 30 dB in the stopband. Also, determine the corresponding polyphase filter structure for implementing the interpolator.

8 Design a sample rate converter that reduces the sampling rate by a factor 25 . Use the Remez algorithm to determine the coefficients of the FIR filter that has a 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Specify the sets of time-variant coefficients g(n, m) and the corresponding coefficients in the polyphase filter realization of the sample rate converter.

833

Multirate Digital Signal Processing

9

Consider the two different ways of cascading a decimator with an interpolator shown in Fig. P9.

↑I

x(n)

↓D

y1(n)

↑I

y2(n)

(a) ↓D

x(n)

(b)

Figure P9

(a) If D = I , show that the outputs of the two configurations are different. Hence, in general, the two systems are not identical. (b) Show that the two systems are identical if and only if D and I are relatively prime. 10 Prove the equivalence of the two decimator and interpolator configurations shown in Fig. P10 (noble identities) (see Vaidyanathan, 1990). x(n)

↓D

y(n) H(z)

x(n) ––

H(z D)

↓D

↑I

H(zI )

y(n)

(a) x(n) H(z)

↑I

y(n)

x(n) ––

y(n)

(b)

Figure P10

11

Consider an arbitrary digital filter with transfer function

H (z) =

∞ 

h(n)z−n

n=−∞

(a) Perform a two-component polyphase decomposition of H (z) by grouping the even-numbered samples h0 (n) = h(2n) and the odd-numbered samples h1 (n) = h(2n + 1). Thus show that H (z) can be expressed as H (z) = H0 (z2 ) + z−1 H1 (z2 ) and determine H0 (z) and H1 (z).

834

Multirate Digital Signal Processing

(b) Generalize the result in part (a) by showing that H (z) can be decomposed into a D -component polyphase filter structure with transfer function H (z) =

D−1 

z−k Hk (zD )

k=0

Determine Hk (z). (c) For the IIR filter with transfer function H (z) =

1 1 − az−1

determine H0 (z) and H1 (z) for the two-component decomposition. A sequence x(n) is upsampled by I = 2, it passes through an LTI system H1 (z), and then it is downsampled by D = 2. Can we replace this process with a single LTI system H2 (z)? If the answer is positive, determine the system function of this system. 13 Plot the signals and their corresponding spectra for rational sampling rate conversion by (a) I /D = 5/3 and (b) I /D = 3/5. Assume that the spectrum of the input signal x(n) occupies the entire range −π ≤ ωx ≤ π. 14 We wish to design an efficient nonrecursive decimator for D = 8 using the factorization K−1 H (z) = [(1 + z−1 )(1 + z−2 )(1 + z−4 ) . . . (1 + z−2 )]5

12

(a) Derive an efficient implementation using filters with system function Hk (z) = (1 + z−1 )5 . (b) Show that each stage of the obtained decimator can be implemented more efficiently using a polyphase decomposition. 15

Let us consider an alternative polyphase decomposition by defining new polyphase filters Qm (zN ) as N−1  H (z) = z−(N −1−m) Qm (zN ) m=0

This polyphase decomposition is called a type II form to distinguish it from the conventional decomposition based on polyphase filters Pm (zN ). (a) Show that the type II form polyphase filters Qm (z) are related to the polyphase filters Pm (zN ) as follows: Qm (zN ) = PN−1−m (zN ) (b) Sketch the polyphase filter structure for H (z) based in the polyphase filters Qm (zN ) and, thus, show that this structure is an alternative transpose form.

835

Multirate Digital Signal Processing

16 17

Use the result in Problem 15 to determine the type II form of the I = 3 interpolator in Figure 5.12(b). Design a two-stage decimator for the following specifications: D = 100 0 ≤ F ≤ 50

Passband:

50 ≤ F ≤ 55

Transition band: Input sampling rate:

10,000 Hz δ1 = 10−1 , δ2 = 10−3

Ripple: 18

Design a linear-phase FIR filter that satisfies the following specifications based on a single-stage and a two-stage multirate structure: Sampling rate:

10,000 Hz 0 ≤ F ≤ 60

Passband:

60 ≤ F ≤ 65

Transition band:

δ1 = 10−1 , δ2 = 10−3

Ripple: 19 20

Prove that the half-band filter that satisfies (11.35) is always odd and the even coefficients are zero. Design one-stage and two-stage interpolators to meet the following specifications: I = 20 Input sampling rate: Passband: Transition band: Ripple:

10,000 Hz 0 ≤ F ≤ 90 90 ≤ F ≤ 100 δ1 = 10−2 , δ2 = 10−3

21

By using (10.15) derive the equations corresponding to the structure for the polyphase synthesis section shown in Fig. 10.6. 22 Show that the transpose of an L-stage interpolator for increasing the sampling rate by an integer factor I is equivalent to an L-stage decimator that decreases the sampling rate by a factor D = I . 23 Sketch the polyphase filter structure for achieving a time advance of (k/I )Tx in a sequence x(n). 24 Prove the following expressions for an interpolator of order I . (a) The impulse response h(n) can be expressed as h(n) =

I −1  k=0

836

pk (n − k)

Multirate Digital Signal Processing

where

 pk (n) =

n = 0, ±I, ±2I, . . . otherwise

pk (n/I ), 0,

(b) H (z) may be expressed as H (z) =

I −1 

z−k pk (z)

k=0

(c) Pk (z) =

∞ I −1 1   h(n)ej 2πl(n−k)/I z−(n−k)/I I n=−∞ l=0

Pk (ω) =

  I −1 1 2π l j (ω−2πl)k/I H ω− e I I l=0

25

Consider the interpolation of a signal by a factor I . The interpolation filter with transfer function H (z) is implemented by a polyphase filter structure based on an alternative (type II) decomposition, namely H (z) =

I −1 

z−(I −1−m) Qm (zI )

m=0

26

Determine and sketch the structure of the interpolator that employs the polyphase filters Qm (z), 0 ≤ m ≤ I − 1. In the polyphase filter structures for a uniform DFT filter bank described in Section 10.1, the polyphase filters in the analysis section were defined by equation (10.10). Instead of this definition, suppose we define the N -band polyphase filters for the lowpass prototype filter, H0 (z), as H0 (z) =

N−1 

z−i Pi (zN )

i=0

where Pi (z) =

∞ 

h0 (nN + i)z−n ,

0≤i ≤N −1

n=0

Then,

Hk (z) = H0 (z e−j 2πk/N ) = H0 (zWNk )

where WN = e−j 2π/N . (a) Show that the filters Hk (z), 1 ≤ k ≤ N − 1, can be expressed as  P0 (zN ) −1   z P1 (zN ) Hk (z) = 1 WN−k WN−2k · · · WN−(N −1)k  ..  .

   

z−(N −1) PN−1 (zN )

837

Multirate Digital Signal Processing

(b) Show that



  H0 (z)  H1 (z)     = N W−1  ..    . HN−1 (z)



P0 (zN ) −1 z P1 (zN )

  ..  . z−(N −1) PN−1 (zN )

where W is the DFT matrix 

1 1 1 WN1   WN2 W = 1 . ..  .. . 1 WN(N −1)

1 WN2 WN4 .. .

WN2(N−1)

··· ··· ···

1



WN(N −1)   WN2(N−1)    ..  . (N −1)(N −1) · · · WN

(c) Sketch the analysis section of the uniform DFT filter bank. (d) Determine and sketch the synthesis section of the uniform DFT filter bank. 27 The prototype filter in a four-channel uniform DFT filter bank is characterized by the transfer function H0 (z) = 1 + z−1 + 3z−2 + 4z−4 (a) Determine the transfer functions of the filters H1 (z), H2 (z) and H3 (z) in the analysis section. (b) Determine the transfer functions of the filters in the synthesis section. (c) Sketch the analysis and synthesis sections of the uniform DFT filter bank. 28

Consider the following FIR filter transfer function: H (z) = −3 + 19z−2 + 32z−3 + 19z−4 − 3z−6 (a) Show that H (z) is a linear-phase filter. (b) Show that H (z) is a half-band filter. (c) Plot the magnitude and phase responses of the filter.

29

The analysis filter H0 (z) in a two-channel QMF has the transfer function H0 (z) = 1 + z−1 (a) Determine the polyphase filters P0 (z2 ) and P1 (z2 ). (b) Determine the analysis filter H1 (z) and sketch the two-channel analysis section that employs polyphase filters. (c) Determine the synthesis filters G0 (z) and G1 (z) and sketch the entire twochannel QMF based on polyphase filters. (d) Show that the QMF bank results in perfect reconstruction.

838

Multirate Digital Signal Processing

30

The analysis filters in a three-channel QMF bank have the transfer functions H0 (z) = 1 + z−1 + z−2 H1 (z) = 1 − z−1 + z−2 H2 (z) = 1 − z−2 (a) Determine the polyphase matrix P(z3 ) and express the analysis filters in the form (12.11). (b) Determine the synthesis filters G0 (z), G1 (z), and G2 (z) that result in perfect reconstruction. (c) Sketch the three-channel QMF bank that employs polyphase filters in the analysis and synthesis sections.

31 Zoom-frequency analysis yR (n)

x(n)

Consider the system in Fig. P31(a).

↓D

LPF

X

e j2πf0n

DFT yI (n)

↓D

LPF

(a) X(ω)

←π

0

ω0 ω0 + ∆ω π

ω

(b)

Figure P31

(a) Sketch the spectrum of the signal y(n) = yR (n) + jyI (n) if the input signal x(n) has the spectrum shown in Fig. P31(b). (b) Suppose that we are interested in the analysis of the frequencies in the band f0 ≤ f ≤ f0 + f , where f0 = π/6 and f = π/3. Determine the cutoff of a lowpass filter and the decimation factor D required to retain the information contained in this band of frequencies.

839

Multirate Digital Signal Processing

(c) Assume that x(n) =

p−1   k=0

 k cos 2πfk n 1− 2p

where p = 40 and fk = k/p, k = 0, 1, . . . , p − 1. Compute and plot the 1024-point DFT of x(n). (d) Repeat part (b) for the signal x(n) given in part (c) by using an appropriately designed lowpass linear phase FIR filter to determine the decimated signal s(n) = sR (n) + j sI (n). (e) Compute the 1024-point DFT of s(n) and investigate to see if you have obtained the expected results.

840

Linear Prediction and Optimum Linear Filters

The design of filters to perform signal estimation is a problem that frequently arises in the design of communication systems, control systems, in geophysics, and in many other applications and disciplines. In this chapter we treat the problem of optimum filter design from a statistical viewpoint. The filters are constrained to be linear and the optimization criterion is based on the minimization of the mean-square error. As a consequence, only the second-order statistics (autocorrelation and crosscorrelation functions) of a stationary process are required in the determination of the optimum filters. Included in this treatment is the design of optimum filters for linear prediction. Linear prediction is a particularly important topic in digital signal processing, with applications in a variety of areas, such as speech signal processing, image processing, and noise suppression in communication systems. As we shall observe, determination of the optimum linear filter for prediction requires the solution of a set of linear equations that have some special symmetry. To solve these linear equations, we describe two algorithms, the Levinson–Durbin algorithm and the Schur algorithm, which provide the solution to the equations through computationally efficient procedures that exploit the symmetry properties. The last section of the chapter treats an important class of optimum filters called Wiener filters. Wiener filters are widely used in many applications involving the estimation of signals corrupted with additive noise.

1

Random Signals, Correlation Functions, and Power Spectra We begin with a brief review of the characterization of random signals in terms of statistical averages expressed in both the time domain and the frequency domain. The

From Chapter 12 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

841

Linear Prediction and Optimum Linear Filters

reader is assumed to have a background in probability theory and random processes, at the level given in the books of Helstrom (1990), Peebles (1987), and Stark and Woods (1994).

1.1

Random Processes

Many physical phenomena encountered in nature are best characterized in statistical terms. For example, meteorological phenomena such as air temperature and air pressure fluctuate randomly as a function of time. Thermal noise voltages generated in the resistors of electronic devices, such as a radio or television receiver, are also randomly fluctuating phenomena. These are just a few examples of random signals. Such signals are usually modeled as infinite-duration infinite-energy signals. Suppose that we take the set of waveforms corresponding to the air temperature in different cities around the world. For each city there is a corresponding waveform that is a function of time, as illustrated in Fig. 1.1. The set of all possible waveforms

Figure 1.1

842

Sample functions of a random process.

Linear Prediction and Optimum Linear Filters

is called an ensemble of time functions or, equivalently, a random process. The waveform for the temperature in any particular city is a single realization or a sample function of the random process. Similarly, the thermal noise voltage generated in a resistor is a single realization or a sample function of the random process consisting of all noise voltage waveforms generated by the set of all resistors. The set (ensemble) of all possible noise waveforms of a random process is denoted as X(t, S), where t represents the time index and S represents the set (sample space) of all possible sample functions. A single waveform in the ensemble is denoted by x(t, s). Usually, we drop the variable s (or S ) for notational convenience, so that the random process is denoted as X(t) and a single realization is denoted as x(t). Having defined a random process X(t) as an ensemble of sample functions, let us consider the values of the process for any set of time instants t1 > t2 > · · · > tn , where n is any positive integer. In general, the samples Xti ≡ x(ti ), i = 1, 2, . . . , n are n random variables characterized statistically by their joint probability density function (PDF) denoted as p(xt1 , xt2 , . . . , xtn ) for any n.

1.2

Stationary Random Processes

Suppose that we have n samples of the random process X(t) at t = ti , i = 1, 2, . . . , n, and another set of n samples displaced in time from the first set by an amount τ . Thus the second set of samples are Xti +τ ≡ X(ti + τ ), i = 1, 2, . . . , n, as shown in Fig. 1. 1. This second set of n random variables is characterized by the joint probability density function p(xti +τ , . . . , xtn +τ ). The joint PDFs of the two sets of random variables may or may not be identical. When they are identical, then p(xt1 , xt2 , . . . , xtn ) = p(xt1 +τ , xt2 +τ , . . . , xtn +τ )

(1.1)

for all τ and all n, and the random process is said to be stationary in the strict sense. In other words, the statistical properties of a stationary random process are invariant to a translation of the time axis. On the other hand, when the joint PDFs are different, the random process is nonstationary.

1.3

Statistical (Ensemble) Averages

Let us consider a random process X(t) sampled at time instant t = ti . Thus X(ti ) is a random variable with PDF p(xti ). The l th moment of the random variable is defined as the expected value of X l (ti ), that is,  ∞ l xtli p(xti )dxti E(Xti ) = (1.2) −∞

In general, the value of the l th moment depends on the time instant ti , if the PDF of Xti depends on ti . When the process is stationary, however, p(xti +τ ) = p(xti ) for all τ . Hence the PDF is independent of time and, consequently, the l th moment is independent of time (a constant).

843

Linear Prediction and Optimum Linear Filters

Next, let us consider the two random variables Xti = X(ti ), i = 1, 2, corresponding to samples of X(t) taken at t = t1 and t = t2 . The statistical (ensemble) correlation between Xt1 and Xt2 is measured by the joint moment  E(Xt1 Xt2 ) =



−∞



∞ −∞

xt1 xt2 p(xt1 xt2 ) dx1 dx2

(1.3)

Since the joint moment depends on the time instants t1 and t2 , it is denoted as γxx (t1 , t2 ) and is called the autocorrelation function of the random process. When the process X(t) is stationary, the joint PDF of the pair (Xt1 , Xt2 ) is identical to the joint PDF of the pair (Xt1 +τ , Xt2 +τ ) for any arbitrary τ . This implies that the autocorrelation function of X(t) depends on the time difference t1 − t2 = τ . Hence for a stationary real-valued random process the autocorrelation function is γxx (τ ) = E[Xt1 +τ Xt1 ]

(1.4)

γxx (−τ ) = E(Xt1 −τ Xt1 ) = E(Xt  Xt  +τ ) = γxx (τ )

(1.5)

On the other hand, 1

1

Therefore, γxx (τ ) is an even function. We also note that γxx (0) = E(Xt21 ) is the average power of the random process. There exist nonstationary processes with the property that the mean value of the process is a constant and the autocorrelation function satisfies the property γxx (t1 , t2 ) = γxx (t1 − t2 ). Such a process is called wide-sense stationary. Clearly, widesense stationarity is a less stringent condition than strict-sense stationarity. In our treatment we shall require only that the processes be wide-sense stationary. Related to the autocorrelation function is the autocovariancefunction, which is defined as cxx (t1 , t2 ) = E{[Xt1 − m(t1 )][Xt2 − m(t2 )]} (1.6) = γxx (t1 , t2 ) − m(t1 )m(t2 ) where m(t1 ) = E(Xt1 ) and m(t2 ) = E(Xt2 ) are the mean values of Xt1 and Xt2 , respectively. When the process is stationary, cxx (t1 , t2 ) = cxx (t1 − t2 ) = cxx (τ ) = γxx (τ ) − m2x

(1.7)

where τ = t1 − t2 . Furthermore, the variance of the process is σx2 = cxx (0) = γxx (0) − m2x .

1.4

Statistical Averages for Joint Random Processes

Let X(t) and Y (t) be two random processes and let Xti ≡ X(ti ), i = 1, 2, . . . , n, and Yt  ≡ Y (tj ), j = 1, 2, . . . , m, represent the random variables at times t1 > t2 > j

844

Linear Prediction and Optimum Linear Filters

· · · > tn and t1 > t2 > · · · > tm , respectively. The two sets of random variables are characterized statistically by the joint PDF p(xt1 , xt2 , . . . , xtn , yt  , yt  , . . . , ytm ) 1

2

for any set of time instants {ti } and {tj } and for any positive integer values of m and n. The crosscorrelation function of X(t) and Y (t), denoted as γxy (t1 , t2 ), is defined by the joint moment  ∞ ∞ xt1 yt2 p(xt1 , yt2 )dxt1 dyt2 (1.8) γxy (t1 , t2 ) ≡ E(Xt1 Yt2 ) = −∞

−∞

and the crosscovariance is cxy (t1 , t2 ) = γxy (t1 , t2 ) − mx (t1 )my (t2 )

(1.9)

When the random processes are jointly and individually stationary, we have γxy (t1 , t2 ) = γxy (t1 − t2 ) and cxy (t1 , t2 ) = cxy (t1 − t2 ). In this case γxy (−τ ) = E(Xt1 Yt1 +τ ) = E(Xt  −τ Yt  ) = γyx (τ ) 1

1

(1.10)

The random processes X(t) and Y (t) are said to be statistically independent if and only if p(xt1 , xt2 , . . . , xtn , yt  , yt  , . . . , ytm ) = p(xt1 , . . . , xtn )p(yt  , . . . , ytm ) 1

2

1

ti

for all choices of ti , and for all positive integers n and m. The processes are said to be uncorrelated if (1.11) γxy (t1 , t2 ) = E(Xt1 )E(Yt2 ) so that cxy (t1 , t2 ) = 0. A complex-valued random process Z(t) is defined as Z(t) = X(t) + j Y (t)

(1.12)

where X(t) and Y (t) are random processes. The joint PDF of the complex-valued random variables Zti ≡ Z(ti ), i = 1, 2, . . . , is given by the joint PDF of the components (Xti , Yti ), i = 1, 2, . . . , n. Thus the PDF that characterizes Zti , i = 1, 2, . . . , n is p(xt1 , xt2 , . . . , xtn , yt1 , yt2 , . . . , ytn ) A complex-valued random process Z(t) is encountered in the representation of the in-phase and quadrature components of the lowpass equivalent of a narrowband random signal or noise. An important characteristic of such a process is its autocorrelation function, which is defined as γzz (t1 , t2 ) = E(Zt1 Zt∗2 ) = E[(Xt1 + j Yt1 )(Xt2 − j Yt2 )]

(1.13)

= γxx (t1 , t2 ) + γyy (t1 , t2 ) + j [γyx (t1 , t2 ) − γxy (t1 , t2 )]

845

Linear Prediction and Optimum Linear Filters

When the random processes X(t) and Y (t) are jointly and individually stationary, the autocorrelation function of Z(t) becomes γzz (t1 , t2 ) = γzz (t1 − t2 ) = γzz (τ ) where τ = t1 − t2 . The complex conjugate of (1.13) is γzz∗ (τ ) = E(Zt∗1 Zt1 −τ ) = γzz (−τ )

(1.14)

Now, suppose that Z(t) = X(t) + j Y (t) and W (t) = U (t) + j V (t) are two complex-valued random processes. Their crosscorrelation function is defined as γzw (t1 , t2 ) = E(Zt1 Wt∗2 ) = E[(Xt1 + j Yt1 )(Ut2 − j Vt2 )]

(1.15)

= γxu (t1 , t2 ) + γyv (t1 , t2 ) + j [γyu (t1 , t2 ) − γxv (t1 , t2 )] When X(t), Y (t), U (t), and V (t) are pairwise stationary, the crosscorrelation functions in (1.15) become functions of the time difference τ = t1 − t2 . In addition, we have ∗ (1.16) γzw (τ ) = E(Zt∗1 Wt1 −τ ) = E(Zt∗1 +τ Wt1 ) = γwz (−τ )

1.5

Power Density Spectrum

A stationary random process is an infinite-energy signal and hence its Fourier transform does not exist. The spectral characteristic of a random process is obtained according to the Wiener–Khintchine theorem, by computing the Fourier transform of the autocorrelation function. That is, the distribution of power with frequency is given by the function  ∞

γxx (τ )e−j 2πF τ dτ

(1.17)

The inverse Fourier transform is given as  ∞ xx (F )ej 2πF τ dF γxx (τ ) =

(1.18)

xx (F ) =

−∞

−∞

We observe that

 γxx (0) = =



−∞

xx (F )dF

E(Xt2 )

(1.19) ≥0

Since E(Xt2 ) = γxx (0) represents the average power of the random process, which is the area under xx (F ), it follows that xx (F ) is the distribution of power as a function of frequency. For this reason, xx (F ) is called the power density spectrum of the random process.

846

Linear Prediction and Optimum Linear Filters

If the random process is real, γxx (τ ) is real and even and hence xx (F ) is real ∗ and even. If the random process is complex valued, γxx (τ ) = γxx (−τ ) and, hence ∗ (F ) = xx



−∞

 =





−∞

∗ γxx (τ )ej 2πF τ dτ =





−∞

∗ γxx (−τ )e−j 2πF τ dτ

γxx (τ )e−j 2πF τ dτ = xx (F )

Therefore, xx (F ) is always real. The definition of the power density spectrum can be extended to two jointly stationary random processes X(t) and Y (t), which have a crosscorrelation function γxy (τ ). The Fourier transform of γxy (τ ) is  xy (F ) =



−∞

γxy (τ )e−j 2πF τ dτ

(1.20)

∗ (F ) = which is called the cross-power density spectrum. It is easily shown that xy yx (−F ). For real random processes, the condition is yx (F ) = xy (−F ).

1.6

Discrete-Time Random Signals

This characterization of continuous-time random signals can be easily carried over to discrete-time signals. Such signals are usually obtained by uniformly sampling a continuous-time random process. A discrete-time random process X(n) consists of an ensemble of sample sequences x(n). The statistical properties of X(n) are similar to the characterization of X(t), with the restriction that n is now an integer (time) variable. To be specific, we state the form for the important moments that we use in this text. The l th moment of X(n) is defined as  ∞ E(Xnl ) = xnl p(xn )dxn (1.21) −∞

and the autocorrelation sequence is  γxx (n, k) = E(Xn Xk ) =



−∞



∞ −∞

xn xk p(xn , xk )dxn dxk

(1.22)

Similarly, the autocovariance is cxx (n, k) = γxx (n, k) − E(Xn )E(Xk )

(1.23)

For a stationary process, we have the special forms (m = n − k) γxx (n − k) = γxx (m) cxx (n − k) = cxx (m) = γxx (m) − m2x

(1.24)

847

Linear Prediction and Optimum Linear Filters

where mx = E(Xn ) is the mean of the random process. The variance is defined as σ 2 = cxx (0) = γxx (0) − m2x . For a complex-valued stationary process Z(n) = X(n) + j Y (n), we have γzz (m) = γxx (m) + γyy (m) + j [γyx (m) − γxy (m)]

(1.25)

and the crosscorrelation sequence of two complex-valued stationary sequences is γzw (m) = γxu (m) + γyv (m) + j [γyu (m) − γxv (m)]

(1.26)

As in the case of a continuous-time random process, a discrete-time random process has infinite energy but a finite average power and is given as E(Xn2 ) = γxx (0)

(1.27)

By use of the Wiener–Khintchine theorem, we obtain the power density spectrum of the discrete-time random process by computing the Fourier transform of the autocorrelation sequence γxx (m), that is, ∞ 

xx (f ) =

γxx (m)e−j 2πf m

(1.28)

xx (f )ej 2πf m df

(1.29)

m=−∞

The inverse transform relationship is  γxx (m) =

1/2

−1/2

We observe that the average power is  γxx (0) =

1/2

−1/2

xx (f )df

(1.30)

so that xx (f ) is the distribution of power as a function of frequency, that is, xx (f ) is the power density spectrum of the random process X(n). The properties we have stated for xx (F ) also hold for xx (f ).

1.7

Time Averages for a Discrete-Time Random Process

Although we have characterized a random process in terms of statistical averages, such as the mean and the autocorrelation sequence, in practice, we usually have available a single realization of the random process. Let us consider the problem of obtaining the averages of the random process from a single realization. To accomplish this, the random process must be ergodic. By definition, a random process X(n) is ergodic if, with probability 1, all the statistical averages can be determined from a single sample function of the process. In effect, the random process is ergodic if time averages obtained from a single

848

Linear Prediction and Optimum Linear Filters

realization are equal to the statistical (ensemble) averages. Under this condition we can attempt to estimate the ensemble averages using time averages from a single realization. To illustrate this point, let us consider the estimation of the mean and the autocorrelation of the random process from a single realization x(n). Since we are interested only in these two moments, we define ergodicity with respect to these parameters. For additional details on the requirements for mean ergodicity and autocorrelation ergodicity which are given below, the reader is referred to the book of Papoulis (1984).

1.8

Mean-Ergodic Process

Given a stationary random process X(n) with mean mx = E(Xn ) let us form the time average m ˆx =

N  1 x(n) 2N + 1

(1.31)

n=−N

In general, we view m ˆ x in (1.31) as an estimate of the statistical mean whose value will vary with the different realizations of the random process. Hence m ˆ x is a random variable with a PDF p(m ˆ x ). Let us compute the expected value of m ˆ x over all possible realizations of X(n). Since the summation and the expectation are linear operations, we can interchange them, so that E(m ˆ x) =

N N   1 1 E[x(n)] = mx = mx 2N + 1 2N + 1 n=−N

(1.32)

n=−N

Since the mean value of the estimate is equal to the statistical mean, we say that the estimate m ˆ x is unbiased. Next, we compute the variance of m ˆ x . We have ˆ x |2 ) − |mx |2 var(m ˆ x ) = E(|m But E(|m ˆ x |2 ) =

N N   1 E[x ∗ (n)x(k)] (2N + 1)2 n=−N k=−N

=

N N   1 γxx (k − n) (2N + 1)2 n=−N k=−N

=

  2N  1 |m| 1− γxx (m) 2N + 1 2N + 1 m=−2N

849

Linear Prediction and Optimum Linear Filters

Therefore, var(m ˆ x) =

  2N  |m| 1 1− γxx − |mx |2 2N + 1 2N + 1 m=−2N

=

1 2N + 1

 2N 



1−

m=−2N

(1.33)

|m| cxx (m) 2N + 1

If var(mx ) → 0 as N → ∞, the estimate converges with probability 1 to the statistical mean mx . Therefore, the process X(n) is mean ergodic if   2N  1 |m| 1− cxx (m) = 0 N→∞ 2N + 1 2N + 1 lim

(1.34)

m=−2N

Under this condition, the estimate m ˆ x in the limit as N → ∞ becomes equal to the statistical mean, that is, N  1 x(n) N→∞ 2N + 1

mx = lim

(1.35)

n=−N

Thus the time-average mean, in the limit as N → ∞, is equal to the ensemble mean. A sufficient condition for (1.34) to hold is if ∞ 

|cxx (m)| < ∞

(1.36)

m=−∞

which implies that cxx (m) → 0 as m → ∞. This condition holds for most zero-mean processes encountered in the physical world.

1.9

Correlation-Ergodic Processes

Now, let us consider the estimate of the autocorrelation γxx (m) from a single realization of the process. Following our previous notation, we denote the estimate (for a complex-valued signal, in general) as N  1 x ∗ (n)x(n + m) rxx (m) = 2N + 1

(1.37)

n=−N

Again, we regard rxx (m) as a random variable for any given lag m, since it is a function of the particular realization. The expected value (mean value over all realizations) is E[rxx (m)] =

N  1 E[x ∗ (n)x(n + m)] 2N + 1 n=−N

=

850

1 2N + 1

N  n=−N

(1.38) γxx (m) = γxx (m)

Linear Prediction and Optimum Linear Filters

Therefore, the expected value of the time-average autocorrelation is equal to the statistical average. Hence we have an unbiased estimate of γxx (m). To determine the variance of the estimate rxx (m), we compute the expected value of |rxx (m)|2 and subtract the square of the mean value. Thus var[rxx (m)] = E[|rxx (m)|2 ] − |γxx (m)|2

(1.39)

But E[|rxx (m)|2 ] =

N N   1 E[x ∗ (n)x(n + m)x(k)x ∗ (k + m)] (2N + 1)2

(1.40)

n=−N k=−N

The expected value of the term x ∗ (n)x(n+m)x(k)x ∗ (k +m) is just the autocorrelation sequence of a random process defined as vm (n) = x ∗ (n)x(n + m) Hence (1.40) may be expressed as

E[|rxx (m)|2 ] =

N N   1 (m) γvv (n − k) (2N + 1)2 n=−N k=−N

=

1 2N + 1

2N   n=−2N



1−

(1.41)

|n| γ (m) (n) 2N + 1 vv

and the variance is  2N   1 |n| 1− γ (m) (n) − |γxx (m)|2 var[rxx (m)] = 2N + 1 2N + 1 vv

(1.42)

n=−2N

If var[rxx (m)] → 0 as N → ∞, the estimate rxx (m) converges with probability 1 to the statistical autocorrelation γxx (m). Under these conditions, the process is correlation ergodic and the time-average correlation is identical to the statistical average, that is, N  1 x ∗ (n)x(n + m) = γxx (m) N→∞ 2N + 1

lim

(1.43)

n=−N

In our treatment of random signals, we assume that the random processes are mean ergodic and correlation ergodic, so that we can deal with time averages of the mean and the autocorrelation obtained from a single realization of the process.

851

Linear Prediction and Optimum Linear Filters

2

Innovations Representation of a Stationary Random Process In this section we demonstrate that a wide-sense stationary random process can be represented as the output of a causal and causally invertible linear system excited by a white noise process. The condition that the system is causally invertible also allows us to represent the wide-sense stationary random process by the output of the inverse system, which is a white noise process. Let us consider a wide-sense stationary process {x(n)} with autocorrelation sequence {γxx (m)} and power spectral density xx (f ), |f | ≤ 21 . We assume that xx (f ) is real and continuous for all |f | ≤ 21 . The z-transform of the autocorrelation sequence {γxx (m)} is ∞  γxx (m)z−m (2.1) xx (z) = m=−∞

from which we obtain the power spectral density by evaluating xx (z) on the unit circle [i.e., by substituting z = exp(j 2πf )]. Now, let us assume that log xx (z) is analytic (possesses derivatives of all orders) in an annular region in the z-plane that includes the unit circle (i.e., r1 < |z| < r2 where r1 < 1 and r2 > 1). Then, log xx (z) can be expanded in a Laurent series of the form ∞  v(m)z−m (2.2) log xx (z) = m=−∞

where the {v(m)} are the coefficients in the series expansion. We can view {v(m)} as the sequence with z-transform V (z) = log xx (z). Equivalently, we can evaluate log xx (z) on the unit circle, ∞ 

log xx (f ) =

v(m)e−j 2πf m

(2.3)

m=−∞

so that the {v(m)} are the Fourier coefficients in the Fourier series expansion of the periodic function log xx (f ). Hence  v(m) =

1 2

− 21

[log xx (f )]ej 2πf m df,

m = 0, ±1, . . .

(2.4)

We observe that v(m) = v(−m), since xx (f ) is a real and even function of f . From (2.2) it follows that  xx (z) = exp

∞ 

 v(m)z

m=−∞

= σw2 H (z)H (z−1 )

852

−m

(2.5)

Linear Prediction and Optimum Linear Filters

where, by definition, σw2 = exp[v(0)] and ∞   H (z) = exp v(m)z−m ,

|z| > r1

(2.6)

m=1

If (2.5) is evaluated on the unit circle, we have the equivalent representation of the power spectral density as xx (f ) = σw2 |H (f )|2

(2.7)

We note that log xx (f ) = log σw2 + log H (f ) + log H ∗ (f ) ∞ 

=

v(m)e−j 2πf m

m=−∞

From the definition of H (z) given by (2.6), it is clear that the causal part of the Fourier series in (2.3) is associated with H (z) and the anticausal part is associated with H (z−1 ). The Fourier series coefficients {v(m)} are the cepstral coefficients and the sequence {v(m)} is called the cepstrum of the sequence {γxx (m)}.. The filter with system function H (z) given by (2.6) is analytic in the region |z| > r1 < 1. Hence, in this region, it has a Taylor series expansion as a causal system of the form ∞  h(n)z−n H (z) = (2.8) m=0

The output of this filter in response to a white noise input sequence w(n) with power spectral density σw2 is a stationary random process {x(n)} with power spectral density xx (f ) = σw2 |H (f )|2 . Conversely, the stationary random process {x(n)} with power spectral density xx (f ) can be transformed into a white noise process by passing {x(n)} through a linear filter with system function 1/H (z). We call this filter a noise whitening filter. Its output, denoted as {w(n)} is called the innovations process associated with the stationary random process {x(n)}. These two relationships are illustrated in Fig. 2.1. w(n) White noise

Linear causal filter H(z)

x(n) =

Σ

h(k)w(n − k)

k=0

(a)

Figure 2.1

Filters for generating (a) the random process x(n) from white noise and (b) the inverse filter.

x(n)

Linear causal filter 1/H(z)

w(n) White noise (b)

853

Linear Prediction and Optimum Linear Filters

The representation of the stationary stochastic process {x(n)} as the output of an IIR filter with system function H (z) given by (2.8) and excited by a white noise sequence {w(n)} is called the Wold representation.

2.1

Rational Power Spectra

Let us now restrict our attention to the case where the power spectral density of the stationary random process {x(n)} is a rational function, expressed as xx (z) = σw2

B(z)B(z−1 ) , A(z)A(z−1 )

r1 < |z| < r2

(2.9)

where the polynomials B(z) and A(z) have roots that fall inside the unit circle in the z-plane. Then the linear filter H (z) for generating the random process {x(n)} from the white noise sequence {w(n)} is also rational and is expressed as q 

B(z) H (z) = = A(z)

bk z−k

k=0 p

1+



|z| > r1 ak z

(2.10)

−k

k=1

where {bk } and {ak } are the filter coefficients that determine the location of the zeros and poles of H (z), respectively. Thus H (z) is causal, stable, and minimum phase. Its reciprocal 1/H (z) is also a causal, stable, and minimum-phase linear system. Therefore, the random process {x(n)} uniquely represents the statistical properties of the innovations process {w(n)}, and vice versa. For the linear system with the rational system function H (z) given by (2.10), the output x(n) is related to the input w(n) by the difference equation x(n) +

p 

ak x(n − k) =

q 

bk w(n − k)

(2.11)

k=0

k=1

We will distinguish among three specific cases. b0 = 1, bk = 0, k > 0. In this case, the linear filter H (z) = 1/A(z) is an all-pole filter and the difference equation for the input–output relationship is p  (2.12) ak x(n − k) = w(n) x(n) +

Autoregressive (AR) process.

k=1

In turn, the noise-whitening filter for generating the innovations process is an all-zero filter.

854

Linear Prediction and Optimum Linear Filters

Moving average (MA) process. ak = 0, k ≥ 1. In this case, the linear filter H (z) = B(z) is an all-zero filter and the difference equation for the input–output relationship is q  bk w(n − k) (2.13) x(n) = k=0

The noise-whitening filter for the MA process is an all-pole filter. Autoregressive, moving average (ARMA) process. H (z) = B(z)/A(z) has both finite poles and zeros sponding difference equation is given by (2.11). erating the innovations process from x(n) is also a 1/H (z) = A(z)/B(z).

In this case, the linear filter in the z-plane and the correThe inverse system for genpole–zero system of the form

2.2

Relationships Between the Filter Parameters and the Autocorrelation Sequence When the power spectral density of the stationary random process is a rational function, there is a basic relationship between the autocorrelation sequence {γxx (m)} and the parameters {ak } and {bk } of the linear filter H (z) that generates the process by filtering the white noise sequence w(n). This relationship can be obtained by multiplying the difference equation in (2.11) by x ∗ (n − m) and taking the expected value of both sides of the resulting equation. Thus we have ∗

E[x(n)x (n − m)] = −

p 

ak E[x(n − k)x ∗ (n − m)]

k=1



(2.14)

q

+

bk E[w(n − k)x ∗ (n − m)]

k=0

Hence γxx (m) = −

p 

ak γxx (m − k) +

q 

bk γwx (m − k)

(2.15)

k=0

k=1

where γwx (m) is the crosscorrelation sequence between w(n) and x(n). The crosscorrelation γwx (m) is related to the filter impulse response. That is, γwx (m) = E[x ∗ (n)w(n + m)] ∞   ∗ =E h(k)w (n − k)w(n + m)

(2.16)

k=0

= σw2 h(−m) where, in the last step, we have used the fact that the sequence w(n) is white. Hence  0, m>0 γwx (m) = (2.17) σw2 h(−m), m ≤ 0

855

Linear Prediction and Optimum Linear Filters

By combining (2.17) with (2.15), we obtain the desired relationship:  p     ak γxx (m − k), m>q −     k=1 q−m p γxx (m) =   2   − a γ (m − k) + σ h(k)bk+m , 0 ≤ m ≤ q k xx  w    k=0  ∗k=1 γxx (−m), m0 −     k=1 p (2.19) γxx (m) =    ak γxx (m − k) + σw2 , m = 0 −      ∗k=1 γxx (−m), mq   ∗ m 0, the zeros |zi | < 1 for every i . The proof is by induction. Clearly, for p = 1 the system function for the prediction-error filter is

A1 (z) = 1 + K1 z−1 f

(5.1)

f

Hence z1 = −K1 and E1 = (1 − |K1 |2 )E0 > 0. Now, suppose that the hypothesis is true for p − 1. Then, if zi is a root of Ap (z), we have from (3.16) and (3.18), Ap (zi ) = Ap−1 (zi ) + Kp zi−1 Bp−1 (zi )   1 −p ∗ = Ap−1 (zi ) + Kp zi Ap−1 =0 zi Hence

(5.2)

−p

∗ zi Ap−1 (1/zi ) 1 =− ≡ Q(zi ) Kp Ap−1 (zi )

(5.3)

We note that the function Q(z) is all pass. In general, an all-pass function of the form P (z) =

N  zz∗ + 1 k

k=1

z + zk

,

|zk | < 1

(5.4)

satisfies the property that |P (z)| > 1 for |z| < 1, |P (z)| = 1 for |z| = 1, and |P (z)| < 1 for |z| > 1. Since Q(z) = −P (z)/z, it follows that |zi | < 1 if |Q(z)| > 1. Clearly, this f is the case since Q(zi ) = 1/Kp and Ep > 0. f f On the other hand, suppose that Ep−1 > 0 and Ep = 0. In this case |Kp | = 1 and |Q(zi )| = 1. Since the MMSE is zero, the random process x(n) is called predictable or deterministic. Specifically, a purely sinusoidal random process of the form x(n) =

M 

αk ej (nωk +θk )

(5.5)

k=1

where the phases {θk } are statistically independent and uniformly distributed over (0, 2π), has the autocorrelation γxx (m) =

M 

αk2 ej mωk

(5.6)

k=1

873

Linear Prediction and Optimum Linear Filters

and the power density spectrum xx (f ) =

M 

αk2 δ(f − fk ),

fk =

k=1

ωk 2π

(5.7)

This process is predictable with a predictor of order p ≥ M . To demonstrate the validity of the statement, consider passing this process through a prediction error filter of order p ≥ M . The MSE at the output of this filter is  Epf =

−1/2

=

xx (f )|Ap (f )|2 df

1/2

M 

−1/2

k=1

 =

1/2

M 

 αk2 δ(f

− fk ) |Ap (f )|2 df

(5.8)

αk2 |Ap (fk )|2

k=1

By choosing M of the p zeros of the prediction-error filter to coincide with the f frequencies {fk }, the MSE Ep can be forced to zero. The remaining p − M zeros can be selected arbitrarily to be anywhere inside the unit circle. Finally, the reader can prove that if a random process consists of a mixture of a continuous power spectral density and a discrete spectrum, the prediction-error filter must have all its roots inside the unit circle. Maximum-phase property of the backward prediction-error filter. function for the backward prediction-error filter of order p is

Bp (z) = z−p Ap∗ (z−1 )

The system

(5.9)

Consequently, the roots of Bp (z) are the reciprocals of the roots of the forward prediction-error filter with system function Ap (z). Hence if Ap (z) is minimum phase, then Bp (z) is maximum phase. However, if the process x(n) is predictable, all the roots of Bp (z) lie on the unit circle. Suppose that the random process x(n) is an AR(p) stationary random process that is generated by passing white noise with variance σw2 through an all-pole filter with system function

Whitening property.

H (z) = 1+

1 p  k=1

874

(5.10) ak z−k

Linear Prediction and Optimum Linear Filters

Then the prediction-error filter of order p has the system function Ap (z) = 1 +

p 

ap (k)z−k

(5.11)

k=1

where the predictor coefficients ap (k) = ak . The response of the prediction-error filter is a white noise sequence {w(n)}. In this case the prediction-error filter whitens the input random process x(n) and is called a whitening filter, as indicated in Section 3.4. More generally, even if the input process x(n) is not an AR process, the predictionerror filter attempts to remove the correlation among the signal samples of the input process. As the order of the predictor is increased, the predictor output x(n) ˆ becomes a closer approximation to x(n) and hence the difference f (n) = x(n) ˆ − x(n) approaches a white noise sequence. The backward prediction errors {gm (k)} from different stages in the FIR lattice filter are orthogonal. That is,

Orthogonality of the backward prediction errors.



E[gm (n)gl∗ (n)]

=

0≤l ≤m−1 l=m

0, b Em ,

(5.12)

This property is easily proved by substituting for gm (n) and gl∗ (n) into (5.12) and carrying out the expectation. Thus E[gm (n)gl∗ (n)] =

=

m 

bm (k)

l 

k=0

j =0

l 

m 

bl∗ (j )

j =0

bl∗ (j )E[x(n − k)x ∗ (n − j )] (5.13) bm (k)γxx (j − k)

k=0

But the normal equations for the backward linear predictor require that m 

 bm (k)γxx (j − k) =

k=0

Therefore, E[gm (n)gl∗ (n)] =

0, b Em ,



j = 1, 2, . . . , m − 1 j =m

(5.14)

f

b = Em , m = l Em 0, 0≤l ≤m−1

(5.15)

Additional properties. There are a number of other interesting properties regarding the forward and backward prediction errors in the FIR lattice filter. These are given here for real-valued data. Their proof is left as an exercise for the reader.

875

Linear Prediction and Optimum Linear Filters

(a) E[fm (n)x(n − i)] = 0,

1≤i≤m

(b) E[gm (n)x(n − i)] = 0,

0≤i ≤m−1

(c) E[fm (n)x(n)] = E[gm (n)x(n − m)] = Em (d) E[fi (n)fj (n)] = Emax (i, j )



(e) E[fi (n)fj (n − t)] = 0, for 

1 ≤ t ≤ i − j, −1 ≥ t ≥ i − j,

0≤t (f) E[gi (n)gj (n − t)] = 0, for 0≥t  Ei , (g) E[fi (n + i)fj (n + j )] = 0,

i>j ij i 0, the linear estimation problem is referred to as signal prediction. Note that this problem is different than the prediction considered earlier in this chapter, where d(n) = x(n + D), D ≥ 0. 3. If d(n) = s(n − D), where D > 0, the linear estimation problem is referred to as signal smoothing. Our treatment will concentrate on filtering and prediction. The criterion selected for optimizing the filter impulse response {h(n)} is the minimization of the mean-square error. This criterion has the advantages of simplicity and mathematical tractability. The basic assumptions are that the sequences {s(n)}, {w(n)}, and {d(n)} are zero mean and wide-sense stationary. The linear filter will be assumed to be either FIR or IIR. If it is IIR, we assume that the input data {x(n)} are available over the infinite past. We begin with the design of the optimum FIR filter. The optimum linear filter, in the sense of minimum mean-square error (MMSE), is called a Wiener filter.

d(n)

s(n) Signal

+

x(n)

Optimum linear filter

+

y(n) −

+

e(n)

Figure 7.1

Model for linear estimation problem.

w(n) Noise

881

Linear Prediction and Optimum Linear Filters

7.1

FIR Wiener Filter

Suppose that the filter is constrained to be of length M with coefficients {hk , 0 ≤ k ≤ M − 1). Hence its output y(n) depends on the finite data record x(n), x(n − 1), . . . , x(n − M + 1), M−1  y(n) = h(k)x(n − k) (7.1) k=0

The mean-square value of the error between the desired output d(n) and y(n) is EM = E|e(n)|2  2 M−1      = E d(n) − h(k)x(n − k)  

(7.2)

k=0

Since this is a quadratic function of the filter coefficients, the minimization of EM yields the set of linear equations M−1 

h(k)γxx (l − k) = γdx (l),

l = 0, 1, . . . , M − 1

(7.3)

k=0

where γxx (k) is the autocorrelation of the input sequence {x(n)} and γdx (k) = E[d(n)x ∗ (n − k)] is the crosscorrelation between the desired sequence {d(n)} and the input sequence {x(n), 0 ≤ n ≤ M − 1}. The set of linear equations that specify the optimum filter is called the Wiener–Hopf equation. These equations are also called the normal equations, encountered earlier in the chapter in the context of linear one-step prediction. In general, the equations in (7.3) can be expressed in matrix form as M hM = γd

(7.4)

where M is an M × M (Hermitian) Toeplitz matrix with elements lk = γxx (l − k) and γd is the M × 1 crosscorrelation vector with elements γdx (l), l = 0, 1, . . . , M − 1. The solution for the optimum filter coefficients is hopt = −1 M γd

(7.5)

and the resulting minimum MSE achieved by the Wiener filter is MMSEM = min EM = σd2 − hM

or, equivalently,

M−1 

∗ hopt (k)γdx (k)



MMSEM = σd2 − γd t −1 M γd

882

(7.6)

k=0

(7.7)

Linear Prediction and Optimum Linear Filters

where σd2 = E|d(n)|2 . Let us consider some special cases of (7.3). If we are dealing with filtering, the d(n) = s(n). Furthermore, if s(n) and w(n) are uncorrelated random sequences, as is usually the case in practice, then γxx (k) = γss (k) + γww (k) γdx (k) = γss (k)

(7.8)

and the normal equations in (7.3) become M−1 

h(k)[γss (l − k) + γww (l − k)] = γss (l),

l = 0, 1, . . . , M − 1

(7.9)

k=0

If we are dealing with prediction, then d(n) = s(n + D) where D > 0. Assuming that s(n) and w(n) are uncorrelated random sequences, we have γdx (k) = γss (l + D)

(7.10)

Hence the equations for the Wiener prediction filter become M−1 

h(k)[γss (l − k) + γww (l − k)] = γss (l + D),

l = 0, 1, . . . , M − 1

(7.11)

k=0

In all these cases, the correlation matrix to be inverted is Toeplitz. Hence the (generalized) Levinson–Durbin algorithm may be used to solve for the optimum filter coefficients. EXAMPLE 7.1 Let us consider a signal x(n) = s(n) + w(n), where s(n) is an AR(1) process that satisfies the difference equation s(n) = 0.6s(n − 1) + v(n) where {v(n)} is a white noise sequence with variance σv2 = 0.64, and {w(n)} is a white noise sequence with variance σw2 = 1. We will design a Wiener filter of length M = 2 to estimate {s(n)}. Solution. Since {s(n)} is obtained by exciting a single-pole filter by white noise, the power spectral density of s(n) is ss (f ) = σv2 |H (f )|2 =

0.64 |1 − 0.6e−j 2πf |2

=

0.64 1.36 − 1.2 cos 2πf

883

Linear Prediction and Optimum Linear Filters

The corresponding autocorrelation sequence {γss (m)} is γss (m) = (0.6)|m| The equations for the filter coefficients are 2h(0) + 0.6h(1) = 1 0.6h(0) + 2h(1) = 0.6 Solution of these equations yields the result h(0) = 0.451,

h(1) = 0.165

The corresponding minimum MSE is MMSE2 = 1 − h(0)γss (0) − h(1)γss (1) = 1 − 0.451 − (0.165)(0.6) = 0.45 This error can be reduced further by increasing the length of the Wiener filter (see Problem 35).

7.2

Orthogonality Principle in Linear Mean-Square Estimation

The normal equations for the optimum filter coefficients given by (7.3) can be obtained directly by applying the orthogonality principle in linear mean-square estimation. Simply stated, the mean-square error EM in (7.2) is a minimum if the filter coefficients {h(k)} are selected such that the error is orthogonal to each of the data points in the estimate, E[e(n)x ∗ (n − l)] = 0, where e(n) = d(n) −

M−1 

l = 0, 1, . . . , M − 1

(7.12)

h(k)x(n − k)

(7.13)

k=0

Conversely, if the filter coefficients satisfy (7.12), the resulting MSE is a minimum. When viewed geometrically, the output of the filter, which is the estimate ˆ d(n) =

M−1 

h(k)x(n − k)

(7.14)

k=0

is a vector in the subspace spanned by the data {x(k), 0 ≤ k ≤ M − 1}. The error ˆ e(n) is a vector from d(n) to dˆ (n) [i.e., d(n) = e(n) + d(n)], as shown in Fig. 7.2. The orthogonality principle states that the length EM = E|e(n)|2 is a minimum when e(n) is perpendicular to the data subspace [i.e., e(n) is orthogonal to each data point x(k), 0 ≤ k ≤ M − 1].

884

Linear Prediction and Optimum Linear Filters

d(n)

e(n) h(0)x(1)

x(1)

h(1)x(2) ˆ d(n)

Figure 7.2

Geometric interpretation of linear MSE problem.

x(2)

We note that the solution obtained from the normal equations in (7.3) is ˆ unique if the data {x(n)} in the estimate d(n) are linearly independent. In this case, the correlation matrix M is nonsingular. On the other hand, if the data are linearly dependent, the rank of M is less than M and therefore the solution is not unique. ˆ In this case, the estimate d(n) can be expressed as a linear combination of a reduced set of linearly independent data points equal to the rank of M . Since the MSE is minimized by selecting the filter coefficients to satisfy the orthogonality principle, the residual minimum MSE is simply MMSEM = E[e(n)d ∗ (n)]

(7.15)

which yields the result given in (7.6).

7.3

IIR Wiener Filter

In the preceding section we constrained the filter to be FIR and obtained a set of M linear equations for the optimum filter coefficients. In this section we allow the filter to be infinite in duration (IIR) and the data sequence to be infinite as well. Hence the filter output is ∞  y(n) = h(k)x(n − k) (7.16) k=0

The filter coefficients are selected to minimize the mean-square error between the desired output d(n) and y(n), that is, E∞ = E|e(n)|2  2 ∞      = E d(n) − h(k)x(n − k)  

(7.17)

k=0

885

Linear Prediction and Optimum Linear Filters

Application of the orthogonality principle leads to the Wiener–Hopf equation, ∞ 

h(k)γxx (l − k) = γdx (l),

l≥0

(7.18)

k=0

The residual MMSE is simply obtained by application of the condition given by (7.15). Thus we obtain MMSE∞ = min E∞ = σd2 − h

∞ 

∗ hopt (k)γdx (k)

(7.19)

k=0

The Wiener–Hopf equation given by (7.18) cannot be solved directly with z-transform techniques, because the equation holds only for l ≥ 0. We shall solve for the optimum IIR Wiener filter based on the innovations representation of the stationary random process {x(n)}. Recall that a stationary random process {x(n)} with autocorrelation γxx (k) and power spectral density xx (f ) can be represented by an equivalent innovations process, {i(n)}, by passing {x(n)} through a noise-whitening filter with system function 1/G(z), where G(z) is the minimum-phase part obtained from the spectral factorization of xx (z): xx (z) = σi2 G(z)G(z−1 ) (7.20) Hence G(z) is analytic in the region |z| > r1 , where r1 < 1. Now, the optimum Wiener filter can be viewed as the cascade of the whitening filter 1/G(z) with a second filter, say Q(z), whose output y(n) is identical to the output of the optimum Wiener filter. Since y(n) =

∞ 

q(k)i(n − k)

(7.21)

k=0

and e(n) = d(n) − y(n), application of the orthogonality principle yields the new Wiener–Hopf equation as ∞ 

q(k)γii (l − k) = γdi (l),

l≥0

(7.22)

k=0

But since {i(n)} is white, it follows that γii (l − k) = 0 unless l = k. Thus we obtain the solution as γdi (l) γdi (l) q(l) = = , l≥0 (7.23) γii (0) σi2 The z-transform of the sequence {q(l)} is Q(z) =

∞ 

q(k)z−k

k=0 ∞ 1  = 2 γdi (k)z−k σi k=0

886

(7.24)

Linear Prediction and Optimum Linear Filters

If we denote the z-transform of the two-sided crosscorrelation sequence γdi (k) by di (z) ∞  γdi (k)z−k (7.25) di (z) = k=−∞

and define [di (z)]+ as

∞ 

[di (z)]+ =

γdi (k)z−k

(7.26)

k=0

then Q(z) =

1 [di (z)]+ σi2

(7.27)

To determine [di (z)]+ , we begin with the output of the noise-whitening filter, which can be expressed as i(n) =

∞ 

v(k)x(n − k)

(7.28)

k=0

where {v(k), k ≥ 0} is the impulse response of the noise-whitening filter, ∞

 1 ≡ V (z) = v(k)z−k G(z)

(7.29)

k=0

Then γdi (k) = E[d(n)i ∗ (n − k)] =

∞ 

v(m)E[d(n)x ∗ (n − m − k)]

(7.30)

m=0

=

∞ 

v(m)γdx (k + m)

m=0

The z-transform of the crosscorrelation γdi (k) is ∞  ∞   v(m)γdx (k + m) z−k di (z) = k=−∞

=

∞  m=0

=

∞  m=0

m=0

v(m)

∞ 

γdx (k + m)z−k

k=−∞

v(m)zm

(7.31)

∞ 

γdx (k)z−k

k=−∞

= V (z−1 )dx (z) =

dx (z) G(z−1 )

887

Linear Prediction and Optimum Linear Filters

Therefore, Q(z) =

1 σi2



dx (z) G(z−1 )

 (7.32) +

Finally, the optimum IIR Wiener filter has the system function Hopt (z) = =

Q(z) G(z)

  1 dx (z) σi2 G(z) G(z−1 ) +

(7.33)

In summary, the solution for the optimum IIR Wiener filter requires that we perform a spectral factorization of xx (z) to obtain G(z), the minimum-phase component, and then we solve for the causal part of dx (z)/G(z−1 ). The following example illustrates the procedure. EXAMPLE 7.2 Let us determine the optimum IIR Wiener filter for the signal given in Example 7.1. Solution.

For this signal we have xx (z) = ss (z) + 1 =

where σi2 = 1.8 and G(z) =

1.8(1 − 13 z−1 )(1 − 13 z) (1 − 0.6z−1 )(1 − 0.6z)

1 − 13 z−1 1 − 0.6z−1

The z-transform of the crosscorrelation γdx (m) is dx (z) = ss (z) = Hence



dx (z) G(z−1 )



 +

=

=

0.64



(1 − 13 z)(1 − 0.6z−1 ) 

=

0.64 (1 − 0.6z−1 )(1 − 0.6z)

0.8 0.266z + −1 1 − 0.6z 1 − 13 z



+

+

0.8 1 − 0.6z−1

The optimum IIR filter has the system function    1 − 0.6z−1 1 0.8 Hopt (z) = 1.8 1 − 13 z−1 1 − 0.6z−1 =

888

1

4 9 − 13 z−1

Linear Prediction and Optimum Linear Filters

and an impulse response hopt (n) =

4 9

 1 n 3

,

n≥0

We conclude this section by expressing the minimum MSE given by (7.19) in terms of the frequency-domain characteristics of the filter. First, we note that σd2 ≡ E|d(n)|2 is simply the value of the autocorrelation sequence {γdd (k)} evaluated at k = 0. Since  1 dd (z)zk−1 dz γdd (k) = (7.34) 2πj Cˆ it follows that σd2

 1 dd (z) = γdd (0) = dz ˆ 2πj C z

(7.35)

where the contour integral is evaluated along a closed path encircling the origin in the region of convergence of dd (z). The second term in (7.19) is also easily transformed to the frequency domain by application of Parseval’s theorem. Since hopt (k) = 0 for k < 0, we have ∞ 

∗ hopt (k)γdx (k) =

k=−∞

 1 Hopt (z)dx (z−1 )z−1 dz 2πj Cˆ

(7.36)

where C is a closed contour encircling the origin that lies within the common region of convergence of Hopt (z) and dx (z−1 ). By combining (7.35) with (7.36), we obtain the desired expression for the MMSE ∞ in the form MMSE∞

 1 = [dd (z) − Hopt (z)dx (z−1 )]z−1 dz 2πj Cˆ

(7.37)

EXAMPLE 7.3 For the optimum Wiener filter derived in Example 7.2, the minimum MSE is

MMSE∞

   1 0.3555 = dz 2πj Cˆ (z − 13 )(1 − 0.6z)

There is a single pole inside the unit circle at z = 13 . By evaluating the residue at the pole, we obtain MMSE∞ = 0.444 We observe that this MMSE is only slightly smaller than that for the optimum two-tap Wiener filter in Example 7.1.

889

Linear Prediction and Optimum Linear Filters

7.4

Noncausal Wiener Filter

In the preceding section we constrained the optimum Wiener filter to be causal [i.e., hopt (n) = 0 for n < 0]. In this section we drop this condition and allow the filter to include both the infinite past and the infinite future of the sequence {x(n)} in forming the output y(n), that is, ∞  h(k)x(n − k) (7.38) y(n) = k=−∞

The resulting filter is physically unrealizable. It can also be viewed as a smoothing ˆ filter in which the infinite future signal values are used to smooth the estimate d(n) = y(n) of the desired signal d(n). Application of the orthogonality principle yields the Wiener–Hopf equation for the noncausal filter in the form ∞ 

h(k)γxx (l − k) = γdx (l),

−∞ < l < ∞

(7.39)

k=−∞

and the resulting MMSE nc as MMSEnc = σd2 −

∞ 

∗ h(k)γdx (k)

(7.40)

k=−∞

Since (7.39) holds for −∞ < l < ∞, this equation can be transformed directly to yield the optimum noncausal Wiener filter as Hnc (z) =

dx (z) xx (z)

The MMSE nc can also be simply expressed in the z-domain as  1 [dd (z) − Hnc (z)dx (z−1 )]z−1 dz MMSEnc = 2πj Cˆ

(7.41)

(7.42)

In the following example we compare the form of the optimal noncausal filter with the optimal causal filter obtained in the previous section. EXAMPLE 7.4 The optimum noncausal Wiener filter for the signal characteristics given in Example 7.1 is given by (7.41), where dx (z) = ss (z) = and

xx (z) = ss (z) + 1 =

890

0.64 (1 − 0.6z−1 )(1 − 0.6z)

2(1 − 0.3z−1 − 0.3z) (1 − 0.6z−1 )(1 − 0.6z)

Linear Prediction and Optimum Linear Filters

Then, Hnc (z) =

0.3555 (1 − 13 z−1 )(1 − 13 z)

This filter is clearly noncausal. The minimum MSE achieved by this filter is determined from evaluating (7.42). The integrand is 0.3555 1 ss (z)[1 − Hnc (z)] = z (z − 13 )(1 − 13 z) The only pole inside the unit circle is z = 13 . Hence the residue is  0.3555   1 − 1z 3

= z= 13

0.3555 = 0.40 8/9

Hence the minimum achievable MSE obtained with the optimum noncausal Wiener filter is MMSEnc = 0.40 Note that this is lower than the MMSE for the causal filter, as expected.

8

Summary and References The major focal point in this chapter is the design of optimum linear systems for linear prediction and filtering. The criterion for optimality is the minimization of the mean-square error between a specified desired filter output and the actual filter output. In the development of linear prediction, we demonstrated that the equations for the forward and backward prediction errors specified a lattice filter whose parameters, the reflection coefficients {Km }, were simply related to the filter coefficients {am (k)} of the direct-form FIR linear predictor and the associated prediction-error filter. The optimum filter coefficients {Km } and {am (k)} are easily obtained from the solution of the normal equations. We described two computationally efficient algorithms for solving the normal equations, the Levinson–Durbin algorithm and the Schur algorithm. Both algorithms are suitable for solving a Toeplitz system of linear equations and have a computational complexity of O(p2 ) when executed on a single processor. However, with full parallel processing, the Schur algorithm solves the normal equations in O(p) time, whereas the Levinson–Durbin algorithm requires O(p log p) time. In addition to the all-zero lattice filter resulting from linear prediction, we also derived the AR lattice (all-pole) filter structure and the ARMA lattice-ladder (pole– zero) filter structure. Finally, we described the design of the class of optimum linear filters, called Wiener filters. Linear estimation theory has had a long and rich history of development over the past four decades. Kailath (1974) presents a historical account of the first three

891

Linear Prediction and Optimum Linear Filters

decades. The pioneering work of Wiener (1949) on optimum linear filtering for statistically stationary signals is especially significant. The generalization of the Wiener filter theory to dynamical systems with random inputs was developed by Kalman (1960) and Kalman and Bucy (1961). Kalman filters are treated in the books by Meditch (1969), Brown (1983), and Chui and Chen (1987). The monograph by Kailath (1981) treats both Wiener and Kalman filters. There are numerous references on linear prediction and lattice filters. Tutorial treatments on these subjects have been published in the journal papers by Makhoul (1975, 1978) and Friedlander (1982a, b). The books by Haykin (1991), Markel and Gray 1976), and Tretter (1976) provide comprehensive treatments of these subjects. Applications of linear prediction to spectral analysis are found in the books by Kay (1988) and Marple (1987), to geophysics in the book by Robinson and Treitel (1980), and to adaptive filtering in the book by Haykin (1991). The Levinson–Durbin algorithm for solving the normal equations recursively was given by Levinson (1947) and later modified by Durbin (1959). Variations of this classical algorithm, called split Levinson algorithms, have been developed by Delsarte and Genin (1986) and by Krishna (1988). These algorithms exploit additional symmetries in the Toeplitz correlation matrix and save about a factor of 2 in the number of multiplications. The Schur algorithm was originally described by Schur (1917) in a paper published in German. An English translation of this paper appears in the book edited by Gohberg (1986). The Schur algorithm is intimately related to the polynomials {Am (z)}, which can be interpreted as orthogonal polynomials. A treatment of orthogonal polynomials is given in the books by Szegö (1967), Grenander and Szegö (1958), and Geronimus (1958). The thesis of Vieira (1977) and the papers by Kailath et al. (1978), Delsarte et al. (1978), and Youla and Kazanjian (1978) provide additional results on orthogonal polynomials. Kailath (1985, 1986) provides tutorial treatments of the Schur algorithm and its relationship to orthogonal polynomials and the Levinson–Durbin algorithm. The pipelined parallel processing structure for computing the reflection coefficients based on the Schur algorithm and the related problem of solving Toeplitz systems of linear equations is described in the paper by Kung and Hu (1983). Finally, we should mention that some additional computational efficiency can be achieved in the Schur algorithm, by further exploiting symmetry properties of Toeplitz matrices, as described by Krishna (1988). This leads to the so-called split-Schur algorithm, which is analogous to the split-Levinson algorithm.

Problems 1

The power density spectrum of an AR process {x(n)} is given as xx (ω) = =

σw2 |A(ω)|2 25 |1 −

e−j ω

+ 21 e−j 2ω |2

where σw2 is the variance of the input sequence.

892

Linear Prediction and Optimum Linear Filters

(a) Determine the difference equation for generating the AR process when the excitation is white noise. (b) Determine the system function for the whitening filter. 2

An ARMA process has an autocorrelation {γxx (m)} whose z-transform is given as xx (z) = 9

(z − 13 )(z − 3) (z −

1 2 )(z

− 2)

,

1 2

< |z| < 2

(a) Determine the filter H (z) for generating {x(n)} from a white noise input sequence. Is H (z) unique? Explain. (b) Determine a stable linear whitening filter for the sequence {x(n)}. 3 Consider the ARMA process generated by the difference equation x(n) = 1.6x(n − 1) − 0.63x(n − 2) + w(n) + 0.9w(n − 1) (a) Determine the system function of the whitening filter and its poles and zeros. (b) Determine the power density spectrum of {x(n)}. 4 Determine the lattice coefficients corresponding to the FIR filter with system function H (z) = A3 (z) = 1 + 5

13 −1 24 z

+ 58 z−2 + 13 z−3

Determine the reflection coefficients {Km } of the lattice filter corresponding to the FIR filter described by the system function H (z) = A2 (z) = 1 + 2z−1 + 31 z−2

6

(a) Determine the zeros and sketch the zero pattern for the FIR lattice filter with reflection coefficients K1 = 21 ,

K2 = − 13 ,

K3 = 1

(b) Repeat part (a) but with K3 = −1. (c) You should have found that the zeros lie on the unit circle. Can this result be generalized? How? 7 Determine the impulse response of the FIR filter that is described by the lattice coefficients K1 = 0.6, K2 = 0.3, K3 = 0.5, and K4 = 0.9. 8 In Section 3.4 we indicated that the noise-whitening filter A p (z) for a causal AR(p) process is a forward linear prediction-error filter of order p. Show that the backward linear prediction-error filter of order p is the noise-whitening filter of the corresponding anticausal AR(p) process. 9 Use the orthogonality principle to determine the normal equations and the resulting minimum MSE for a forward predictor of order p that predicts m samples (m > 1) into the future (m-step forward predictor). Sketch the prediction-error filter. 10 Repeat Problem 9 for an m-step backward predictor.

893

Linear Prediction and Optimum Linear Filters

11

Determine a Levinson–Durbin recursive algorithm for solving for the coefficients of a backward prediction-error filter. Use the result to show that coefficients of the forward and backward predictors can be expressed recursively as     am−1 bm−1 + Km am = 0 1     bm−1 ∗ am−1 bm = + Km 0 1

12

The Levinson–Durbin algorithm described in Section 4.1 solved the linear equations m am = −γm where the right-hand side of this equation has elements of the autocorrelation sequence that are also elements of the matrix . Let us consider the more general problem of solving the linear equations m bm = cm where cm is an arbitrary vector. (The vector bm is not related to the coefficients of the backward predictor.) Show that the solution to m bm = cm can be obtained from a generalized Levinson–Durbin algorithm which is given recursively as bm (m) =

bt bm−1 c(m) − γm−1 f

Em−1

k = 1, 2, . . . , m − 1 m = 1, 2, . . . , p

∗ bm (k) = bm−1 (k) − bm (m)am−1 (m − k), f

13 14

where b1 (1) = c(1)/γ xx (0) = c(1)/E0 and am (k) is given by (4.17). Thus a second recursion is required to solve the equation m bm = cm . Use the generalized Levinson–Durbin algorithm to solve the normal equations recursively for the m-step forward and backward predictors. Consider the AR(3) process generated by the equation x(n) =

14 24 x(n

− 1) +

9 24 x(n

− 2) −

1 24 x(n

− 3) + w(n)

where w(n) is a stationary white noise process with variance σw2 . (a) Determine the coefficients of the optimum p = 3 linear predictor. (b) Determine the autocorrelation sequence γxx (m), 0 ≤ m ≤ 5. (c) Determine the reflection coefficients corresponding to the p = 3 linear predictor. 15 The z-transform of the autocorrelation γxx (m) of an ARMA(1, 1) process is xx (z) = σw2 H (z)H (z−1 ) xx (z) =

894

4σw2 5 − 2z − 2z−1 9 10 − 3z−1 − 3z

Linear Prediction and Optimum Linear Filters

(a) Determine the minimum-phase system function H (z). (b) Determine the system function H (z) for a mixed-phase stable system. 16 Consider an FIR filter with coefficient vector [1

−2r cos θ

r2 ]

(a) Determine the reflection coefficients for the corresponding FIR lattice filter. (b) Determine the values of the reflection coefficients in the limit as r → 1. 17 An AR(3) process is characterized by the prediction coefficients a3 (1) = −1.25,

a3 (2) = 1.25,

a3 (3) = −1

(a) Determine the reflection coefficients. (b) Determine γxx (m) for 0 ≤ m ≤ 3. (c) Determine the mean-square prediction error. 18 The autocorrelation sequence for a random process is  1,     −0.5,

m=0 m = ±1 γxx (m) = 0.625, m = ±2     −0.6875, m = ±3 0 otherwise

19

Determine the system functions Am (z) for the prediction-error filters for m = 1, 2, 3, the reflection coefficients {Km }, and the corresponding mean-square prediction errors. The autocorrelation sequence for an AR process x(n) is γxx (m) = ( 41 )|m|

(a) Determine the difference equation for x(n). (b) Is your answer unique? If not, give any other possible solutions. 20 Repeat Problem 19 for an AR process with autocorrelation γxx (m) = a |m| cos

21

πm 2

where 0 < a < 1. Prove that a FIR filter with system function Ap (z) = 1 +

p 

ap (k)z−k

k=1

and reflection coefficients |Kk | < 1 for 1 ≤ k ≤ p − 1 and |Kp | > 1 is maximum phase [all the roots of Ap (z) lie outside the unit circle].

895

Linear Prediction and Optimum Linear Filters

22 Show that the transformation  Vm =

1 Km∗

Km 1



in the Schur algorithm satisfies the special property t Vm JVm = (1 − |Km |2 )J

where

 J=

23 24 25

26

1 0

0 −1



Thus Vm is called a J -rotation matrix. Its role is to rotate or hyperbolate the row of Gm to lie along the first coordinate direction (Kailath, 1985). Prove the additional properties (a) through (l) of the prediction-error filters given in Section 5. Extend the additional properties (a) through (l) of the prediction-error filters given in Section 5 to complex-valued signals. Determine the reflection coefficient K3 in terms of the autocorrelations {γxx (m)} from the Schur algorithm and compare your result with the expression for K3 obtained from the Levinson–Durbin algorithm. Consider an infinite-length (p = ∞) one-step forward predictor for a stationary random process {x(n)} with a power density spectrum of xx (f ). Show that the mean-square error of the prediction-error filter can be expressed as  1/2 f = 2π exp{ ln xx (f ) df } E∞ −1/2

27

Determine the output of an infinite-length (p = ∞) m-step forward predictor and the resulting mean-square error when the input signal is a first-order autoregressive process of the form x(n) = ax(n − 1) + w(n)

28 An AR(3) process {x(n)} is characterized by the autocorrelation sequence γxx (0) = 1 1, γxx (1) = 21 , γxx (2) = 81 , and γxx (3) = 64 . (a) Use the Schur algorithm to determine the three reflection coefficients K1 , K2 , and K3 . (b) Sketch the lattice filter for synthesizing {x(n)} from a white noise excitation. 29 The purpose of this problem is to show that the polynomials {Am (z)}, which are the system functions of the forward prediction-error filters of order m, m = 0, 1, . . . , p, can be interpreted as orthogonal on the unit circle. Toward this end, suppose that xx (f ) is the power spectral density of a zero-mean random process {x(n)} and let {Am (z)}, m = 0, 1, . . . , p}, be the system functions of the corresponding predictionerror filters. Show that the polynomials {Am (z)} satisfy the orthogonality property  1/2 f xx (f )Am (f )A∗n (f )df = Em δmn , m, n = 0, 1, . . . , p −1/2

896

Linear Prediction and Optimum Linear Filters

30

Determine the system function of the all-pole filter described by the lattice coefficients K1 = 0.6, K2 = 0.3, K3 = 0.5, and K4 = 0.9. 31 Determine the parameters and sketch the lattice-ladder filter structure for the system with system function 1 − 0.8z−1 + 0.15z−2 H (z) = 1 + 0.1z−1 − 0.72z−2 32

Consider a signal x(n) = s(n) + w(n), where s(n) is an AR(1) process that satisfies the difference equation s(n) = 0.8s(n − 1) + v(n) where {v(n)} is a white noise sequence with variance σv2 = 0.49 and {w(n)} is a white noise sequence with variance σw2 = 1. The processes {v(n)} and {w(n)} are uncorrelated. (a) Determine the autocorrelation sequences {γss (m)} and {γxx (m)}. (b) Design a Wiener filter of length M = 2 to estimate {s(n)}.

33 34 35

36

(c) Determine the MMSE for M = 2. Determine the optimum causal IIR Wiener filter for the signal given in Problem 32 and the corresponding MMSE ∞ . Determine the system function for the noncausal IIR Wiener filter for the signal given in Problem 32 and the corresponding M MSEnc . Determine the optimum FIR Wiener filter of length M = 3 for the signal in Example 7.1 and the corresponding MMSE 3 . Compare MMSE 3 with MMSE 2 and comment on the difference. An AR(2) process is defined by the difference equation x(n) = x(n − 1) − 0.6x(n − 2) + w(n)

where {w(n)} is a white noise process with variance σw2 . Use the Yule–Walker equations to solve for the values of the autocorrelation γxx (0), γxx (1), and γxx (2). 37 An observed random process {x(n)} consists of the sum of an AR(p) process of the form p  ap (k)s(n − k) + v(n) s(n) = − k=1

and a white noise process {w(n)} with variance σw2 . The random process {v(n)} is also white with variance σv2 . The sequences {v(n)} and {w(n)} are uncorrelated. Show that the observed process {x(n) = s(n) + w(n)} is ARMA(p, p) and determine the coefficients of the numerator polynomial (MA component) in the corresponding system function.

897

This page intentionally left blank

Adaptive Filters

From Chapter 13 of Digital Signal Processing: Principles, Algorithms, and Applications, Fourth Edition. John G. Proakis, Dimitris G. Manolakis. Copyright © 2007 by Pearson Education, Inc. All rights reserved.

899

Adaptive Filters

In contrast to filter design techniques based on knowledge of the second-order statistics of the signals, there are many digital signal processing applications in which these statistics cannot be specified a priori. Such applications include channel equalization, echo cancellation, and system modeling among others, as described in this chapter. In these applications, filters with adjustable coefficients, called adaptive filters, are employed. Such filters incorporate algorithms that allow the filter coefficients to adapt to the signal statistics. Adaptive filters have received considerable attention from researchers over the past 25 years. As a result, many computationally efficient algorithms for adaptive filtering have been developed. In this chapter, we describe two basic algorithms: the least-mean-square (LMS) algorithm, which is based on a gradient optimization for determining the coefficients, and the class of recursive least-squares algorithms, which includes both direct-form FIR (finite-duration impulse response) and lattice realizations. Before we describe the algorithms, we present several practical applications in which adaptive filters have been successfully used in the estimation of signals corrupted by noise and other interference.

1

Applications of Adaptive Filters Adaptive filters have been used widely in communication systems, control systems, and various other systems in which the statistical characteristics of the signals to be filtered are either unknown a priori or, in some cases, are slowly time-variant (nonstationary signals). Numerous applications of adaptive filters have been described in the literature. Some of the more noteworthy applications include: (1) adaptive antenna systems, in which adaptive filters are used for beam steering and for providing

900

Adaptive Filters

nulls in the beam pattern to remove undesired interference (Widrow, Mantey, and Griffiths (1967)); (2) digital communication receivers, in which adaptive filters are used to provide equalization of intersymbol interference and for channel identification (Lucky (1965), Proakis and Miller (1969), Gersho (1969), George, Bowen, and Storey (1971), Proakis (1970; 1975), Magee and Proakis (1973), Picinbono (1978) and Nichols, Giordano, and Proakis (1977)); (3) adaptive noise cancelling techniques, in which adaptive filters are used to estimate and eliminate a noise component in a desired signal (Widrow et al. (1975), Hsu and Giordano (1978), and Ketchum and Proakis (1982)); (4) system modeling, in which adaptive filters are used as models to estimate the characteristics of an unknown system. These are just a few of the best-known examples of the use of adaptive filters. Although both IIR (infinite-duration impulse response) and FIR filters have been considered for adaptive filtering, the FIR filter is by far the most practical and widely used. The reason for this preference is quite simple: the FIR filter has only adjustable zeros; hence, it is free of stability problems associated with adaptive IIR filters, which have adjustable poles as well as zeros. We should not conclude, however, that adaptive FIR filters are always stable. On the contrary, the stability of the filter depends critically on the algorithm for adjusting its coefficients, as will be demonstrated in Sections 2 and 3. Of the various FIR filter structures that are possible, the direct form and the lattice form are the ones used in adaptive filtering applications. The direct-form FIR filter structure with adjustable coefficients h(n) is illustrated in Figure 1.1. On the other hand, the adjustable parameters in an FIR lattice structure are the reflection coefficients Kn . An important consideration in the use of an adaptive filter is the criterion for optimizing the adjustable filter parameters. The criterion must not only provide a meaningful measure of filter performance, but it must also result in a practically realizable algorithm. For example, a desirable performance index in a digital communication system is the average probability of error. Consequently, in implementing an adaptive equalizer, we might consider the selection of the equalizer coefficients to minimize the average probability of error as the basis for our optimization criterion. Unfortunately, Input

h(0)

z1

z1

z1

h(1)

h(2)

h(3)

z1

h(4)



Output

Coefficient adjustment

Figure 1.1 Direct-form adaptive FIR filter.

901

Adaptive Filters

however, the performance index (average probability of error) for this criterion is a highly nonlinear function of the filter coefficients and the signal statistics. As a consequence, the implementation of an adaptive filter that optimizes such a performance index is complex and impractical. In some cases, a performance index that is a nonlinear function of the filter parameters possesses many relative minima (or maxima), so that one is not certain whether the adaptive filter has converged to the optimum solution or to one of the relative minima (or maxima). For such reasons, some desirable performance indices, such as the average probability of error in a digital communication system, must be rejected on the grounds that they are impractical to implement. Two criteria that provide good measures of performance in adaptive filtering applications are the least-squares criterion and its counterpart in a statistical formulation of the problem, namely, the mean-square-error (MSE) criterion. The leastsquares (and MSE) criterion results in a quadratic performance index as a function of the filter coefficients and, hence, it possesses a single minimum. The resulting algorithms for adjusting the coefficients of the filter are relatively easy to implement, as we demonstrate in Sections 2 and 3. In the following section, we describe several applications of adaptive filters that serve as a motivation for the mathematical development of algorithms derived in Sections 2 and 3. We find it convenient to use the direct-form FIR structure in these examples. Although we will not develop the recursive algorithms for automatically adjusting the filter coefficients in this section, it is instructive to formulate the optimization of the filter coefficients as a least-squares optimization problem. This development will serve to establish a common framework for the algorithms derived in the next two sections.

1.1

System Identification or System Modeling

In the formulation of this problem we have an unknown system, called a plant, that we wish to identify. The system is modeled by an FIR filter with M adjustable coefficients. Both the plant and model are excited by an input sequence x(n). If y(n) denotes the output of the plant and y(n) ˆ denotes the output of the model, y(n) ˆ =

M−1 

h(k)x(n − k)

(1.1)

k=0

We may form the error sequence e(n) = y(n) − y(n), ˆ

n = 0, 1, . . .

(1.2)

and select the coefficients h(k) to minimize ᏱM =

N  n=0

 y(n) −

M−1  k=0

where N + 1 is the number of observations.

902

2 h(k)x(n − k)

(1.3)

Adaptive Filters

Noise w(n) Unknown time-variant system

d(n) y(n) 

x(n)  FIR filter model

ˆ y(n)

Figure 1.2

Application of adaptive filtering to system identification.

Adaptive algorithm

Error signal

The least-squares criterion leads to the set of linear equations for determining the filter coefficients, namely, M−1 

h(k)rxx (l − k) = ryx (l),

l = 0, 1, . . . , M − 1

(1.4)

k=0

In (1.4), rxx (l) is the autocorrelation of the sequence x(n) and ryx (l) is the crosscorrelation of the system output with the input sequence. By solving (1.4), we obtain the filter coefficients for the model. Since the filter parameters are obtained directly from measurement data at the input and output of the system, without prior knowledge of the plant, we call the FIR filter model an adaptive filter. If our only objective were to identify the system by use of the FIR model, the solution of (1.4) would suffice. In control systems applications, however, the system being modeled may be time variant, changing slowly with time, and our purpose for having a model is to ultimately use it for designing a controller that controls the plant. Furthermore, measurement noise is usually present at the output of the plant. This noise introduces uncertainty in the measurements and corrupts the estimates of the filter coefficients in the model. Such a scenario is illustrated in Figure 1.2. In this case, the adaptive filter must identify and track the time-variant characteristics of the plant in the presence of measurement noise at the output of the plant. The algorithms to be described in Sections 2 and 3 are applicable to this system identification problem.

1.2

Adaptive Channel Equalization

Figure 1.3 shows a block diagram of a digital communication system in which an adaptive equalizer is used to compensate for the distortion caused by the transmission medium (channel). The digital sequence of information symbols a(n) is fed to the transmitting filter, whose output is s(t) =

∞ 

a(k)p(t − kTs )

(1.5)

k=0

903

Adaptive Filters

a(n) Data sequence

Channel (time-variant filter)

Transmitter (filter)

Receiver (filter)

Sampler

Noise

Reference signal

aˆ(n)

Decision device

aˆ(n)



d(n)

Adaptive equalizer

 Adaptive algorithm

Error signal

Figure 1.3 Application of adaptive filtering to adaptive channel equalization.

where p(t) is the impulse response of the filter at the transmitter and Ts is the time interval between information symbols; that is, 1/Ts is the symbol rate. For purposes of this discussion, we may assume that a(n) is a multilevel sequence that takes on values from the set ±1, ±3, ±5, . . . , ±(K − 1), where K is the number of possible symbol values. Typically, the pulse p(t) is designed to have the characteristics illustrated in Figure 1.4. Note that p(t) has amplitude p(0) = 1 at t = 0 and p(nT ) = s 0 at t = nTs , n = ±1, ±2, . . . . As a consequence, successive pulses transmitted sequentially every Ts seconds do not interfere with one another when sampled at the time instants t = nTs . Thus, a(n) = s(nTs ). The channel, which is usually well modeled as a linear filter, distorts the pulse and, thus, causes intersymbol interference. For example, in telephone channels, filters are used throughout the system to separate signals in different frequency ranges. These filters cause phase and amplitude distortion. Figure 1.5 illustrates the effect of channel distortion on the pulse p(t) as it might appear at the output of a telephone channel. Now, we observe that the samples taken every Ts seconds are corrupted by p(t) 1

0 5Ts 4Ts 3Ts 2Ts

Ts

t Ts

2Ts

3Ts

4Ts

5Ts

Figure 1.4 Pulse shape for digital transmission of symbols at a rate of 1/Ts symbols per second.

904

Adaptive Filters

q(t)

5Ts 4Ts 3Ts 2Ts Ts

0

t Ts

2Ts

3Ts

4Ts

5Ts

Figure 1.5 Effect of channel distortion on the signal pulse in Figure 1.4.

interference from several adjacent symbols. The distorted signal is also corrupted by additive noise, which is usually wideband. At the receiving end of the communication system, the signal is first passed through a filter that is designed primarily to eliminate the noise outside of the frequency band occupied by the signal. We may assume that this filter is a linear-phase FIR filter that limits the bandwidth of the noise but causes negligible additional distortion on the channel-corrupted signal. Samples of the received signal at the output of this filter reflect the presence of intersymbol interference and additive noise. If we ignore for the moment the possible time variations in the channel, we may express the sampled output at the receiver as

x(nTs ) =

∞ 

a(k)q(nTs − kTs ) + w(nTs )

k=0

= a(n)q(0) +

∞ 

a(k)q(nTs − kTs ) + w(nTs )

(1.6)

k=0 k=n

where w(t) represents the additive noise and q(t) represents the distorted pulse at the output of the receiver filter. To simplify our discussion, we assume that the sample q(0) is normalized to unity by means of an automatic gain control (AGC) contained in the receiver. Then, the

905

Adaptive Filters

sampled signal given in (1.6) may be expressed as ∞ 

x(n) = a(n) +

a(k)q(n − k) + w(n)

(1.7)

k=0 k=n

where x(n) ≡ x(nTs ), q(n) ≡ q(nTs ), and w(n) ≡ w(nTs ). The term a(n) in (1.7) is the desired symbol at the nth sampling instant. The second term, ∞ 

a(k)q(n − k)

k=0 k=n

constitutes the intersymbol interference due to the channel distortion, and w(n) represents the additive noise in the system. In general, the channel distortion effects embodied through the sampled values q(n) are unknown at the receiver. Furthermore, the channel may vary slowly with time such that the intersymbol interference effects are time variant. The purpose of the adaptive equalizer is to compensate the signal for the channel distortion, so that the resulting signal can be detected reliably. Let us assume that the equalizer is an FIR filter with M adjustable coefficients h(n). Its output may be expressed as a(n) ˆ =

M−1 

h(k)x(n + D − k)

(1.8)

k=0

where D is some nominal delay in processing the signal through the filter and a(n) ˆ represents an estimate of the nth information symbol. Initially, the equalizer is trained by transmitting a known data sequence d(n). Then, the equalizer output, a(n), ˆ is compared with d(n) and an error is generated that is used to optimize the filter coefficients. If we again adopt the least-squares error criterion, we select the coefficients h(k) to minimize the quantity

ᏱM =

N  n=0

[d(n) − a(n)] ˆ = 2

N 

 d(n) −

n=0

M−1 

2 h(k)x(n + D − k)

(1.9)

k=0

The result of the optimization is a set of linear equations of the form M−1 

h(k)rxx (l − k) = rdx (l − D),

l = 0, 1, 2, . . . , M − 1

(1.10)

k=0

where rxx (l) is the autocorrelation of the sequence x(n) and rdx (l) is the crosscorrelation between the desired sequence d(n) and the received sequence x(n).

906

Adaptive Filters

Although the solution of (1.10) is obtained recursively in practice (as demonstrated in the following two sections), in principle, we observe that these equations result in values of the coefficients for the initial adjustment of the equalizer. After the short training period, which usually lasts less than one second for most channels, the transmitter begins to transmit the information sequence a(n). In order to track the possible time variations in the channel, the equalizer coefficients must continue to be adjusted in an adaptive manner while receiving data. As illustrated in Figure 1.3, this is usually accomplished by treating the decisions at the output of the decision device as correct, and using the decisions in place of the reference d(n) to generate the error signal. This approach works quite well when decision errors occur infrequently (for example, less than one decision error per hundred symbols). The occasional decision errors cause only small misadjustments in the equalizer coefficients. In Sections 2 and 3, we describe the adaptive algorithms for recursively adjusting the equalizer coefficients.

1.3

Echo Cancellation in Data Transmission over Telephone Channels

In the transmission of data over telephone channels, modems (modulator/demodulators) are used to provide an interface between the digital data sequence and the analog channel. Shown in Figure 1.6 is a block diagram of a communication system in which two terminals, labeled A and B, transmit data by using modems A and B to interface to a telephone channel. As shown, a digital sequence a(n) is transmitted from terminal A to terminal B while another digital sequence b(n) is transmitted from terminal B to A. This simultaneous transmission in both directions is called full-duplex transmission. As described, the two transmitted signals may be represented as sA (t) =

∞ 

a(k)p(t − kTs )

(1.11)

b(k)p(t − kTs )

(1.12)

k=0

sB (t) =

∞  k=0

a(n) Data terminal A

ˆ b(n)

Transmitter A

Receiver B

ˆ a(n)

Receiver A

Transmitter B

b(n)

Modem A

Four-wire telephone channel

Data terminal B

Modem B

Figure 1.6 Full-duplex data transmission over telephone channels.

907

Adaptive Filters

where p(t) is a pulse as shown in Figure 1.4. When a subscriber leases a private line from a telephone company for the purpose of transmitting data between terminals A and B, the telephone line provided is a fourwire line, which is equivalent to having two dedicated telephone (two-wire) channels, one (pair of wires) for transmitting data in one direction and one (pair of wires) for receiving data from the other direction. In such a case the two transmission paths are isolated and, consequently, there is no “crosstalk” or mutual interference between the two signal paths. Channel distortion is compensated by use of an adaptive equalizer, as previously described, at the receiver of each modem. The major problem with the system shown in Figure 1.6 is the cost of leasing a four-wire telephone channel. If the volume of traffic is high and the telephone channel is used either continuously or frequently, as in banking transactions systems or airline reservation systems, the system pictured in Figure 1.6 may be cost effective. Otherwise, it will not be. An alternative solution for low-volume, infrequent transmission of data is to use the dial-up switched telephone network. In this case, the local communication link between the subscriber and the local central telephone office is a two-wire line, called the local loop. At the central office, the subscriber two-wire line is connected to the main four-wire telephone channels that interconnect different central offices, called trunk lines, by a device called a hybrid. By using transformer coupling, the hybrid is tuned to provide isolation between the transmission and reception channels in fullduplex operation. However, due to impedance mismatch between the hybrid and the telephone channel, the level of isolation is often insufficient and, consequently, some of the signal on the transmitter side leaks back and corrupts the signal on the receiver side, causing an “echo” that is often heard in voice communications over telephone channels. To mitigate the echoes in voice transmissions, the telephone companies employ a device called an echo suppressor. In data transmission, the solution is to use an echo canceller within each modem. The echo cancellers are implemented as adaptive filters with automatically adjustable coefficients, just as in the case of transversal equalizers. With the use of hybrids to couple a two-wire to a four-wire channel, and echo cancellers at each modem to estimate and subtract the echoes, the data communication system for the dial-up switched network takes the form shown in Figure 1.7. A hybrid is needed at each modem to isolate the transmitter from the receiver and to couple to the two-wire local loop. Hybrid A is physically located at the central office of subscriber A while hybrid B is located at the central office to which subscriber B is connected. The two central offices are connected by a four-wire line, one pair used for transmission from A to B and the other pair used for transmission in the reverse direction, from B to A. An echo at terminal A due to the hybrid A is called a near-end echo, while an echo at terminal A due to the hybrid B is termed a far-end echo. Both types of echoes are usually present in data transmission and must be removed by the echo canceller.

908

Input data

Transmitter A

Transmitter B

Echo canceller

Echo canceller

Adaptive algorithm

Output data

Receiver A

Modem A

Local loop

Hybrid A

Hybrid B

 

Local loop

Hybrid Adaptive algorithm

 Telephone channel

Adaptive Filters

Hybrid

Input data



Receiver B

Modem B

Figure 1.7 Block diagram model of a digital communication system that uses echo cancellers in the modems.

Output data

909

Adaptive Filters

Suppose we neglect the channel distortion for purposes of this discussion, and deal with the echoes only. The signal received at modem A may be expressed as sRA (t) = A1 sB (t) + A2 sA (t − d1 ) + A3 sA (t − d2 )

(1.13)

where sB (t) is the desired signal to be demodulated at modem A; sA (t −d1 ) is the nearend echo due to hybrid A, sA (t −d2 ) is the far-end echo due to hybrid B; Ai , i = 1, 2, 3, are the corresponding amplitudes of the three signal components; and d1 and d2 are the delays associated with the echo components. A further disturbance that corrupts the received signal is additive noise, so that the received signal at modem A is rA (t) = sRA (t) + w(t)

(1.14)

where w(t) represents the additive noise process. The adaptive echo canceller attempts to estimate adaptively the two echo components. If its coefficients are h(n), n = 0, 1, . . . , M − 1, its output is sˆA (n) =

M−1 

h(k)a(n − k)

(1.15)

k=0

which is an estimate of the echo signal components. This estimate is subtracted from the sampled received signal, and the resulting error signal can be minimized in the least-squares sense to optimally adjust the coefficients of the echo canceller. There are several possible configurations for placement of the echo canceller in the modem, and for forming the corresponding error signal. Figure 1.8 illustrates one configuration, in which the canceller output is subtracted from the sampled output of the receiver filter with input rA (t). Figure 1.9 illustrates a second configuration, in which the echo canceller is generating samples at the Nyquist rate instead of the symbol rate; in this case the error signal used to adjust the coefficients is simply the difference between rA (n), the sampled received signal, and the canceller output. Finally, Figure 1.10 illustrates the canceller operating in combination with an adaptive equalizer. Input data

a(n)

Transmitter filter

Echo canceller Hybrid

Adaptive algorithm  ˆ b(n)



Decision device

 

Symbol rate sampler

Figure 1.8 Symbol rate echo canceller.

910

Receiver filter rA(t)

Adaptive Filters

a(n)

Transmitter filter Echo canceller Hybrid Adaptive algorithm Error signal

ˆ b(n)

Decision device

Receiver filter





Nyquist rate sampler

Figure 1.9 Nyquist rate echo canceller.

Application of the least-squares criterion in any of the configurations shown in Figures 1.8–1.10 leads to a set of linear equations for the coefficients of the echo canceller. The reader is encouraged to derive the equations corresponding to the three configurations.

1.4

Suppression of Narrowband Interference in a Wideband Signal

We now discuss a problem that arises in practice, especially in signal detection and in digital communications. Let us assume that we have a signal sequence v(n) that Input data

a(n)

Transmitter filter

Echo canceller Hybrid

Adaptive algorithm

ˆ b(n)

Decision device 

 

Equalizer

Sampler

Receiver filter



Error signal

Adaptive algorithm

Figure 1.10 Modem with adaptive equalizer and echo canceller.

911

Adaptive Filters

|V( f )|  |X( f )|  |W( f )| |X( f )|

|W( f )|

Figure 1.11

Strong narrowband interference X(f ) in a wideband signal W (f ).

0

fw

f

consists of a desired wideband signal sequence w(n) corrupted by an additive narrowband interference sequence x(n). The two sequences are uncorrelated. These sequences result from sampling an analog signal v(t) at the Nyquist rate (or faster) of the wideband signal w(t). Figure 1.11 illustrates the spectral characteristics of w(n) and x(n). Usually, the interference |X(f )| is much larger than |W (f )| within the narrow frequency band that it occupies. In digital communications and signal detection problems that fit the above model, the desired signal sequence w(n) is often a spread-spectrum signal, while the narrowband interference represents a signal from another user of the frequency band, or intentional interference from a jammer who is trying to disrupt the communications or detection system. Our objective from a filtering viewpoint is to employ a filter that suppresses the narrowband interference. In effect, such a filter will have a notch in the frequency band occupied by |X(f )|, and in practice, the band occupied by |X(f )| is unknown. Moreover, if the interference is nonstationary, its frequency band occupancy may vary with time. Hence, an adaptive filter is desired. From another viewpoint, the narrowband characteristics of the interference allow us to estimate x(n) from past samples of the sequence v(n) and to subtract the estimate from v(n). Since the bandwidth of x(n) is narrow compared to the bandwidth of the sequence w(n), the samples x(n) are highly correlated due to the high sampling rate. On the other hand, the samples w(n) are not highly correlated, since the samples are taken at the Nyquist rate of w(n). By exploiting the high correlation between x(n) and past samples of the sequence v(n), it is possible to obtain an estimate of x(n), which can be subtracted from v(n). The general configuration is illustrated in Figure 1.12. The signal v(n) is delayed by D samples, where D is selected sufficiently large so that the wideband signal components w(n) and w(n − D) contained in v(n) and v(n − D), respectively, are uncorrelated. Usually, a choice of D = 1 or 2 is adequate. The delayed signal sequence v(n−D) is passed through an FIR filter, which is best characterized as a linear predictor of the value x(n) based on M samples v(n − D − k), k = 0, 1, . . . , M − 1. The output of the linear predictor is

912

Adaptive Filters

␷(n)  w(n)  x(n)

e(n)  wˆ (n)

 

␷(nD)

FIR linear predictor

zD Decorrelation delay

Error signal

xˆ (n)

Adaptive algorithm (a)

h(0)

z1

z1

h(1)

h(2)

z1



␷(nD)

z1

h(M  1)

xˆ (n) (b)

Figure 1.12 Adaptive filter for estimating and suppressing a narrowband interference in a wideband signal.

x(n) ˆ =

M−1 

h(k)v(n − D − k)

(1.16)

k=0

This predicted value of x(n) is subtracted from v(n) to yield an estimate of w(n), as illustrated in Figure 1.12. Clearly, the quality of the estimate xˆ (n) determines how well the narrowband interference is suppressed. It is also apparent that the delay D must be kept as small as possible in order to obtain a good estimate of x(n), but must be sufficiently large so that w(n) and w(n − D) are uncorrelated. Let us define the error sequence e(n) = v(n) − x(n) ˆ = v(n) −

M−1 

h(k)v(n − D − k)

(1.17)

k=0

If we apply the least-squares criterion to optimally select the prediction coefficients,

913

Adaptive Filters

we obtain the set of linear equations M−1 

h(k)rvv (l − k) = rvv (l + D),

l = 0, 1, . . . , M − 1

(1.18)

k=0

where rvv (l) is the autocorrelation sequence of v(n). Note, however, that the righthand side of (1.18) may be expressed as rvv (l + D) =

N 

v(n)v(n − l − D)

n=0

=

N 

[w(n) + x(n)][w(n − l − D) + x(n − l − D)]

n=0

= rww (l + D) + rxx (l + D) + rwx (l + D) + rxw (l + D)

(1.19)

The correlations in (1.19) are time-average correlation sequences. The expected value of rww (l + D) is E[rww (l + D)] = 0,

l = 0, 1, . . . , M − 1

(1.20)

because w(n) is wideband and D is large enough that w(n) and w(n − D) are uncorrelated. Also, (1.21) E[rxw (l + D)] = E[rwx (l + D)] = 0 by assumption. Finally, E[rxx (l + D)] = γxx (l + D)

(1.22)

Therefore, the expected value of rvv (l + D) is simply the statistical autocorrelation of the narrowband signal x(n). Furthermore, if the wideband signal is weak relative to the interference , the autocorrelation rvv (l) in the left-hand side of (1.18) is approximately rxx (l). The major influence of w(n) is to the diagonal elements of rvv (l). Consequently, the values of the filter coefficients determined from the linear equations in (1.18) are a function of the statistical characteristics of the interference x(n). The overall filter structure in Figure 1.12 is an adaptive FIR prediction-error filter with coefficients  1, k=0  (1.23) h (k) = −h(k − D), k = D, D + 1, . . . , D + M − 1 0, otherwise and a frequency response H (ω) =

M−1  k=0

914

h (k + D)e−j ωk

(1.24)

Adaptive Filters

0.00

5.00

Filter response (dB)

10.00 15.00 20.00 25.00 30.00 35.00 0.00

0.06

0.12

0.19 0.25 0.31 Frequency (cycles/sample)

0.37

0.44

0.50

f

Figure 1.13 Frequency response characteristics of an adaptive notch filter.

This overall filter acts as a notch filter for the interference. For example, Figure 1.13 illustrates the magnitude of the frequency response of an adaptive filter with M = 15 coefficients, which attempts to suppress a narrowband interference that occupies 20 percent of the frequency band of a desired spread-spectrum signal sequence. The data were generated pseudorandomly by adding a narrowband interference consisting of 100 randomly phased, equal-amplitude sinusoids to a pseudonoise spread-spectrum signal. The coefficients of the filter were obtained by solving the equations in (1.18), with D = 1, where the correlation rvv (l) was obtained from the data. We observe that the overall interference suppression filter has the characteristics of a notch filter. The depth of the notch depends on the power of the interference relative to the wideband signal. The stronger the interference, the deeper the notch. The algorithms presented in Sections 2 and 3 are appropriate for estimating the predictor coefficients continuously, in order to track a nonstationary narrowband interference signal.

1.5

Adaptive Line Enhancer

In the preceding example, the adaptive linear predictor was used to estimate the narrowband interference for the purpose of suppressing the interference from the input sequence v(n). An adaptive line enhancer (ALE) has the same configuration as the interference suppression filter in Figure 1.12, except that the objective is different. In the adaptive line enhancer, x(n) is the desired signal and w(n) represents

915

Adaptive Filters

a wideband noise component that masks x(n). The desired signal x(n) is either a spectral line or a relatively narrowband signal. The linear predictor shown in Figure 1.12(b) operates in exactly the same fashion as that in Figure 1.12(a), and provides an estimate of the narrowband signal x(n). It is apparent that the ALE(i.e., the FIR prediction filter) is a self-tuning filter that has a peak in its frequency res ponse at the frequency of the sinusoid or, equivalently, in the frequency band of the narrowband signal x(n). By having a narrow bandwidth, the noise w(n) outside of the band is suppressed and, thus, the spectral line is enhanced in amplitude relative to the noise power in w(n). This explains why the FIR predictor is called an ALE. Its coefficients are determined by the solution of (1.18).

1.6

Adaptive Noise Cancelling

Echo cancellation, the suppression of narrowband interference in a wideband signal, and the ALE are related to another form of adaptive filtering called adaptive noise cancelling. A model for the adaptive noise canceller is illustrated in Figure 1.14. The primary input signal consists of a desired signal sequence x(n) corrupted by an additive noise sequence w1 (n) and an additive interference (noise) w2 (n). The additive interference (noise) is also observable after it has been filtered by some unknown linear system that yields v2 (n) and is further corrupted by an additive noise sequence w3 (n). Thus, we have available a secondary signal sequence, which may be expressed as v(n) = v2 (n) + w3 (n). The sequences w1 (n), w2 (n), and w3 (n) are assumed to be mutually uncorrelated and zero mean. As shown in Figure 1.14, an adaptive FIR filter is used to estimate the interference sequence w2 (n) from the secondary signal v(n) and subtract the estimate w2 (n) from the primary signal. The output sequence, which represents an estimate of the desired signal x(n), is the error signal e(n) = y(n) − w ˆ 2 (n) = y(n) −

M−1 

h(k)v(n − k)

(1.25)

k=0

x(n)  w1(n)



Observed signal y(n)

Output

 w2(n)

Observed signal ␷(n) Unknown linear system H(z)

␷2(n)

Adaptive FIR filter

w3(n)

Figure 1.14 Example of an adaptive noise-cancelling system.

916

wˆ 2(n)

Error signal

Adaptive algorithm

Adaptive Filters

This error sequence is used to adaptively adjust the coefficients of the FIR filter. If the least-squares criterion is used to determine the filter coefficients, the result of the optimization is the set of linear equations M−1 

h(k)rvv (l − k) = ryv (l),

l = 0, 1, . . . , M − 1

(1.26)

k=0

where rvv (l) is the sample (time-average) autocorrelation of the sequence v(n) and ryv (l) is the sample crosscorrelation of the sequences y(n) and v(n). Clearly, the noise cancelling problem is similar to the last three adaptive filtering applications described above.

1.7

Linear Predictive Coding of Speech Signals

Various methods have been developed over the past four decades for digital encoding of speech signals. In the telephone system, for example, two commonly used methods for speech encoding are pulse code modulation (PCM) and differential PCM (DPCM). These are examples of waveform coding methods. Other waveform coding methods have also been developed, such as delta modulation (DM) and adaptive DPCM. Since the digital speech signal is ultimately transmitted from the source to a destination, a primary objective in devising speech encoders is to minimize the number of bits required to represent the speech signal, while maintaining speech intelligibility. This objective has led to the development of a class of low bit-rate (10,000 bits per second and below) speech encoding methods, which are based on constructing a model of the speech source and transmitting the model parameters. Adaptive filtering finds application in these model-based speech coding systems. We describe one very effective method called linear predictive coding (LPC). In LPC, the vocal tract is modeled as a linear all-pole filter having the system function G (1.27) H (z) = p 1 − k=1 ak z−k where p is the number of poles, G is the filter gain, and ak are the parameters that determine the poles. There are two mutually exclusive excitation functions, used to model voiced and unvoiced speech sounds. On a short-time basis, voiced speech is periodic with a fundamental frequency F0 , or a pitch period 1/F0 , which depends on the speaker. Thus, voiced speech is generated by exciting the all-pole filter model by a periodic impulse train with a period equal to the desired pitch period. Unvoiced speech sounds are generated by exciting the all-pole filter model by the output of a random-noise generator. This model is shown in Figure 1.15. Given a short-time segment of a speech signal, the speech encoder at the transmitter must determine the proper excitation function, the pitch period for voiced speech, the gain parameter G, and the coefficients {ak }. A block diagram that illustrates the source encoding system is given in Figure 1.16. The parameters of the model are determined adaptively from the data. Then, the speech samples are

917

Adaptive Filters

White noise generator Voiced and unvoiced switch

All-pole filter

Speech signal

Periodic impulse generator

Figure 1.15 Block diagram model for the generation of a speech signal.

synthesized by using the model, and an error signal sequence is generated (as shown in Figure 1.16) by taking the difference between the actual and the synthesized sequence. The error signal and the model parameters are encoded into a binary sequence and transmitted to the destination. At the receiver, the speech signal is synthesized from the model and the error signal. The parameters of the all-pole filter model are easily determined from the speech samples by means of linear prediction. To be specific, consider the system shown in Figure 1.17 and assume that we have N signal samples. The output of the FIR filter is p  ak x(n − k) (1.28) x(n) ˆ = k=1

and the corresponding error between the observed sample x(n) and the estimate x(n) ˆ is p  (1.29) ak x(n − k) e(n) = x(n) − k=1

Sampled



x(n)

output from speech source Determine parameters of model and excitation



e(n) xˆ(n)

E n c o d e r

Parameter {ak, G}

Excitation

All-pole model

xˆ(n)

Figure 1.16 Source encoder for a speech signal.

918

Channel

Adaptive Filters



x(n) Speech samples

Error signal

 xˆ(n)

Adaptive FIR predictor

z1

Adaptive algorithm

Figure 1.17 Estimation of pole parameters in LPC.

By applying the least-squares criterion, we can determine the model parameters ak . The result of this optimization is a set of linear equations p 

ak rxx (l − k) = rxx (l),

l = 1, 2, . . . , p

(1.30)

k=1

where rxx (l) is the time-average autocorrelation of the sequence x(n). The gain parameter for the filter can be obtained by noting that its input–output equation is x(n) =

p 

ak x(n − k) + Gv(n)

(1.31)

k=1

where v(n) is the input sequence. Clearly, Gv(n) = x(n) −

p 

ak x(n − k)

k=1

= e(n) Then, G2

N−1 

v 2 (n) =

n=0

N−1 

e2 (n)

(1.32)

n=0

If the input excitation is normalized to unit energy by design, then G = 2

N−1 

e2 (n)

n=0

= rxx (0) −

p 

ak rxx (k)

(1.33)

k=1

Thus, G2 is set equal to the residual energy resulting from the least-squares optimization.

919

Adaptive Filters

In this development, we have described the use of linear prediction to adaptively determine the pole parameters and the gain of an all-pole filter model for speech generation. In practice, due to the nonstationary character of speech signals, this model is applied to short-time segments (10 to 20 milliseconds) of a speech signal. Usually, a new set of parameters is determined for each short-time segment. However, it is often advantageous to use the model parameters measured from previous segments to smooth out sharp discontinuities that usually exist in estimates of model parameters obtained from segment to segment. Although our discussion was totally in terms of the FIR filter structure, we should mention that speech synthesis is usually performed by using the FIR lattice structure and the reflection coefficients Ki . Since the dynamic range of the Ki is significantly smaller than that of the ak , the reflection coefficients require fewer bits to represent them. Hence, the Ki are transmitted over the channel. Consequently, it is natural to synthesize the speech at the destination using the all-pole lattice structure. In our treatment of LPC for speech coding, we have not considered algorithms for the estimation of the excitation and the pitch period. A discussion of appropriate algorithms for these parameters of the model would take us too far afield and, hence, is omitted. The interested reader is referred to Rabiner and Schafer (1978) and Deller, Hansen and Proakis (2000) for a detailed treatment of speech analysis and synthesis methods.

1.8

Adaptive Arrays

In the previous examples, we considered adaptive filtering performed on a single data sequence. However, adaptive filtering has also been widely applied to multiple data sequences that result from antenna, hydrophone, and seismometer arrays, where the sensors (antennas, hydrophones, or seismometers) are arranged in some spatial configuration. Each element of the array of sensors provides a signal sequence. By properly combining the signals from the various sensors, it is possible to change the directivity pattern of the array. For example, consider a linear antenna array consisting of five elements, as shown in Figure 1.18(a). If the signals are simply linearly summed, we obtain the sequence x(n) =

5 

xk (n)

(1.34)

k=1

which results in the antenna directivity pattern shown in Figure 1.18(a). Now, suppose that an interference signal is received from a direction corresponding to one of the sidelobes in the array. By properly weighting the sequences xk (n) prior to combining, it is possible to alter the sidelobe pattern such that the array contains a null in the direction of the interference, as shown in Figure 1.18(b). Thus, we obtain 5  (1.35) hk xk (n) x(n) = k=1

where the hk are the weights.

920

Adaptive Filters

x1(n)

Side lobes

x2(n) Main lobe To combiner

x3(n)

Look direction

x4(n)

Int

erf

x5(n)

ere

nce

(a)

x1(n)

x2(n) To combiner

x3(n)

Look direction

x4(n)

In

ter

fer

x5(n)

en

ce

(b)

Figure 1.18 Linear antenna array: (a) linear antenna array with antenna pattern;

(b) linear antenna array with a null placed in the direction of the interference.

We may also change or steer the direction of the main antenna lobe by simply introducing delays in the output of the sensor signals prior to combining. Hence, from K sensors we have a combined signal of the form x(n) =

K 

hk xk (n − nk )

(1.36)

k=1

921

Adaptive Filters

where the hk are the weights and nk corresponds to an nk -sample delay in the signal x(n). The choice of weights may be used to place nulls in specific directions. More generally, we may simply filter each sequence prior to combining. In such a case, the output sequence has the general form y(n) =

K 

yk (n)

k=1

=

K M−1  

hk (l)xk (n − nk − l)

(1.37)

k=1 l=0

where hk (l) is the impulse response of the filter for processing the kth sensor output and the nk are the delays that steer the beam pattern. The LMS algorithm described in Section 2.2 is frequently used in adaptively selecting the weights hk or the impulse responses hk (l). The more powerful recursive least-squares algorithms described in Section 3 can also be applied to the multisensor (multichannel) data problem.

2

Adaptive Direct-Form FIR Filters—The LMS Algorithm From the examples of the previous section, we observed that there is a common framework in all the adaptive filtering applications. The least-squares criterion that we have adopted leads to a set of linear equations for the filter coefficients, which may be expressed as M−1 

h(k)rxx (l − k) = rdx (l + D),

l = 0, 1, 2, . . . , M − 1

(2.1)

k=0

where rxx (l) is the autocorrelation of the sequence x(n) and rdx (l) is the crosscorrelation of the sequences d(n) and x(n). The delay parameter D is zero in some cases and nonzero in others. We observe that the autocorrelation rxx (l) and the crosscorrelation rdx (l) are obtained from the data and, hence, represent estimates of the true (statistical) autocorrelation and crosscorrelation sequences. As a result, the coefficients h(k) obtained from (2.1) are estimates of the true coefficients. The quality of the estimates depends on the length of the data record that is available for estimating rxx (l) and rdx (l). This is one problem that must be considered in the implementation of an adaptive filter. A second problem that must be considered is that the underlying random process x(n) is usually nonstationary. For example, in channel equalization, the frequency response characteristics of the channel may vary with time. As a consequence, the statistical autocorrelation and crosscorrelation sequences—and, hence, their estimates—vary with time. This implies that the coefficients of the adaptive filter must change with time to incorporate the time-variant statistical characteristics

922

Adaptive Filters

of the signal into the filter. This also implies that the quality of the estimates cannot be made arbitrarily high by simply increasing the number of signal samples used in the estimation of the autocorrelation and crosscorrelation sequences. There are several ways by which the coefficients of the adaptive filter can be varied with time to track the time-variant statistical characteristics of the signal. The most popular method is to adapt the filter recursively on a sample-by-sample basis, as each new signal sample is received. A second approach is to estimate rxx (l) and rdx (l) on a block-by-block basis, with no attempt to maintain continuity in the values of the filter coefficients from one block of data to another. In such a scheme, the block size must be relatively small, encompassing a time interval that is short compared to the time interval over which the statistical characteristics of the data change significantly. In addition to this block processing method, other block processing schemes can be devised that incorporate some block-to-block continuity in the filter coefficients. In our treatment of adaptive filtering algorithms, we consider only time-recursive algorithms that update the filter coefficients on a sample-by-sample basis. In particular, we consider two types of algorithms, the LMS algorithm, which is based on a gradient-type search for tracking the time-variant signal characteristics, and the class of recursive least-squares algorithms, which are significantly more complex than the LMS algorithm, but which provide faster convergence to changes in signal statistics.

2.1

Minimum Mean-Square-Error Criterion

The LMS algorithm that is described in the following subsection is most easily obtained by formulating the optimization of the FIR filter coefficients as an estimation problem based on the minimization of the mean-square error. Let us assume that we have available the (possibly complex-valued) data sequence x(n), which consists of samples from a stationary random process with autocorrelation sequence γxx (m) = E[x(n)x ∗ (n − m)]

(2.2)

From these samples, we form an estimate of the desired sequence d(n) by passing the observed data x(n) through an FIR filter with coefficients h(n), 0 ≤ n ≤ M − 1. The filter output may be expressed as ˆ d(n) =

M−1 

h(k)x(n − k)

(2.3)

k=0

ˆ where d(n) represents an estimate of d(n). The estimation error is defined as ˆ e(n) = d(n) − d(n) = d(n) −

M−1 

h(k)x(n − k)

(2.4)

k=0

923

Adaptive Filters

The mean-square error as a function of the filter coefficients is ᏱM = E[|e(n)|2 ]  2  M−1      h(k)x(n − k)  = E d(n) −   k=0

 = E |d(n)|2 − 2Re

M−1 

 h∗ (l)d(n)x ∗ (n − l) +

M−1  M−1 

k=0

= σd2 − 2Re

M−1 

h∗ (l)h(k)x ∗ (n − l)x(n − k)

k=0 l=0

 h∗ (l)γdx (l) +

l=0



M−1  M−1 

h∗ (l)h(k)γxx (l − k)

(2.5)

l=0 k=0

where, by definition, σd2 = E[|d(n)|2 ]. We observe that the MSE is a quadratic function of the filter coefficients. Consequently, the minimization of ᏱM with respect to the coefficients leads to the set of M linear equations, M−1 

h(k)γxx (l − k) = γdx (l),

l = 0, 1, . . . , M − 1

(2.6)

k=0

The filter with coefficients obtained from (2.6), which is the Wiener–Hopf equation is called the Wiener filter. If we compare (2.6) with (2.1), it is apparent that these equations are similar in form. In (2.1), we use estimates of the autocorrelation and crosscorrelation to determine the filter coefficients, whereas in (2.6) the statistical autocorrelation and crosscorrelation are employed. Hence, (2.6) yields the optimum (Wiener) filter coefficients in the MSE sense, whereas (2.1) yields estimates of the optimum coefficients. The equations in (2.6) may be expressed in matrix form as M hM = γd

(2.7)

where hM denotes the vector of coefficients, M is an M × M (Hermitian) Toeplitz matrix with elements lk = γxx (l − k), and γd is an M × 1 crosscorrelation vector with elements γdx (l), l = 0, 1, . . . , M − 1. The complex-conjugate of hM is denoted t ∗ as hM and the transpose as hM . The solution for the optimum filter coefficients is hopt = −1 M γd

(2.8)

and the resulting minimum MSE achieved with the optimum coefficients given by (2.8) is ᏱM min = σd2 −

M−1 

∗ hopt (k)γdx (k)

k=0

= σd2 − γdH −1 M γd

924

(2.9)

Adaptive Filters

where the exponent H denotes the conjugate transpose. Recall that the set of linear equations in (2.6) can also be obtained by invoking the orthogonality principle in mean-square estimation. According to the orthogonality principle, the mean-square estimation error is minimized when the error e(n) ˆ is orthogonal, in the statistical sense, to the estimate d(n), that is, E[e(n)dˆ∗ (n)] = 0

(2.10)

But the condition in (2.10) implies that E

M−1 

 ∗

h(k)e(n)x (n − k) =

k=0

M−1 

h(k)E[e(n)x ∗ (n − k)] = 0

k=0

or, equivalently, E[e(n)x ∗ (n − l)] = 0,

l = 0, 1, . . . , M − 1

(2.11)

If we substitute for e(n) in (2.11) using the expression for e(n) given in (2.4), and perform the expectation operation, we obtain the equations given in (2.6). ˆ Since d(n) is orthogonal to e(n), the residual (minimum) mean-square error is ᏱM min = E[e(n)d ∗ (n)] = E[|d(n)|2 ] −

M−1 

∗ hopt (k)γdx (k)

(2.12)

k=0

which is the result given in (2.9). The optimum filter coefficients given by (2.8) can be solved efficiently by using the Levinson-Durbin algorithm. However, we shall consider the use of a gradient method for solving for hopt , iteratively. This development leads to the LMS algorithm for adaptive filtering.

2.2

The LMS Algorithm

There are various numerical methods that can be used to solve the set of linear equations given by (2.6) or (2.7) for the optimum FIR filter coefficients. In the following, we consider recursive methods that have been devised for finding the minimum of a function of several variables. In our problem, the performance index is the MSE given by (2.5), which is a quadratic function of the filter coefficients. Hence, this function has a unique minimum, which we shall determine by an iterative search. For the moment, let us assume that the autocorrelation matrix M and the crosscorrelation vector γd are known. Hence, ᏱM is a known function of the coefficients h(n), 0 ≤ n ≤ M − 1. Algorithms for recursively computing the filter coefficients and, thus, searching for the minimum of ᏱM , have the form 1 hM (n + 1) = hM (n) + (n)S(n), 2

n = 0, 1, . . .

(2.13)

925

Adaptive Filters

where hM (n) is the vector of filter coefficients at the nth iteration, (n) is a step size at the nth iteration, and S(n) is a direction vector for the nth iteration. The initial vector hM (0) is chosen arbitrarily. In this treatment we exclude methods that require the computations of −1 M , such as Newton’s method, and consider only search methods based on the use of gradient vectors. The simplest method for finding the minimum of ᏱM recursively is based on a steepest-descent search (see Murray (1972)). In the method of steepest descent, the direction vector S(n) = −g(n), where g(n) is the gradient vector at the nth iteration, defined as g(n) =

dᏱM (n) dhM (n)

= 2[M hM (n) − γd ],

n = 0, 1, 2, . . .

(2.14)

Hence, we compute the gradient vector at each iteration and change the values of hM (n) in a direction opposite the gradient. Thus, the recursive algorithm based on the method of steepest descent is 1 hM (n + 1) = hM (n) − (n)g(n) 2

(2.15)

hM (n + 1) = [I − (n)M ]hM (n) + (n)γd

(2.16)

or, equivalently,

We state without proof that the algorithm leads to the convergence of hM (n) to hopt in the limit as n → ∞, provided that the sequence of step sizes (n) is absolutely summable, with (n) → 0 as n → ∞. It follows that as n → ∞, g(n) → 0. Other candidate algorithms that provide faster convergence are the conjugategradient algorithm and the Fletcher–Powell algorithm. In the conjugate-gradient algorithm, the direction vectors are given as S(n) = β(n − 1)S(n − 1) − g(n)

(2.17)

where β(n) is a scalar function of the gradient vectors (see Beckman (1960)). In the Fletcher–Powell algorithm, the direction vectors are given as S(n) = −H(n)g(n)

(2.18)

where H(n) is an M ×M positive definite matrix, computed iteratively, that converges to the inverse of M (see Fletcher and Powell (1963)). Clearly, the three algorithms differ in the manner in which the direction vectors are computed. These three algorithms are appropriate when M and γd are known. However, this is not the case in adaptive filtering applications, as we have previously indicated. ˆ In the absence of knowledge of M and γd , we may substitute estimates S(n) of the direction vectors in place of the actual vectors S(n). We consider this approach for the steepest-descent algorithm.

926

Adaptive Filters

First, we note that the gradient vector given by (2.14) may also be expressed in terms of the orthogonality conditions given by (2.11). In fact, the conditions in (2.11) are equivalent to the expression ∗ (n)] = γd − M hM (n) E[e(n)XM

(2.19)

where XM (n) is the vector with elements x(n − l), l = 0, 1, . . . , M − 1. Therefore, the gradient vector is simply ∗ (n)] g(n) = −2E[e(n)XM

(2.20)

Clearly, the gradient vector g(n) = 0 when the error is orthogonal to the data in the ˆ estimate d(n). An unbiased estimate of the gradient vector at the nth iteration is simply obtained from (2.20) as ∗ gˆ (n) = −2e(n)XM (n) (2.21) ˆ where e(n) = d(n) − d(n), and XM (n) is the set of M signal samples in the filter at the nth iteration. Thus, with gˆ (n) substituted for g(n), we have the algorithm ∗ (n) hM (n + 1) = hM (n) + (n)e(n)XM

(2.22)

This is called a stochastic-gradient-descent algorithm. As given by (2.22), it has a variable step size. It has become common practice in adaptive filtering to use a fixed-step-size algorithm for two reasons. The first is that a fixed-step-size algorithm is easily implemented in either hardware or software. The second is that a fixed step size is appropriate for tracking time-variant signal statistics, whereas if (n) → 0 as n → ∞, adaptation to signal variations cannot occur. For these reasons, (2.22) is modified to the algorithm ∗ hM (n + 1) = hM (n) + e(n)XM (n) (2.23) where  is now the fixed step size. This algorithm was first proposed by Widrow and Hoff (1960) and is now widely known as the LMS (least-mean-squares) algorithm. Clearly, it is a stochastic-gradient algorithm. The LMS algorithm is relatively simple to implement. For this reason, it has been widely used in many adaptive filtering applications. Its properties and limitations have also been thoroughly investigated. In the following section, we provide a brief treatment of its important properties concerning convergence, stability, and the noise resulting from the use of estimates of the gradient vectors. Subsequently, we compare its properties with the more complex recursive least-squares algorithms.

2.3

Related Stochastic Gradient Algorithms

Several variations of the basic LMS algorithm have been proposed in the literature and implemented in adaptive filtering applications. One variation is obtained if we

927

Adaptive Filters

average the gradient vectors over several iterations prior to making adjustments of the filter coefficients. For example, the average over K gradient vectors is gˆ (nK) = −

K−1 2  ∗ e(nK + k)XM (nK + k) K

(2.24)

k=0

and the corresponding recursive equation for updating the filter coefficients once every K iterations is 1 hM ((n + 1)K) = hM (nK) − ˆg(nK) 2

(2.25)

In effect, the averaging operation performed in (2.24) reduces the noise in the estimate of the gradient vector, as shown by Gardner (1984). An alternative approach is to filter the gradient vectors by a lowpass filter and use the output of the filter as an estimate of the gradient vector. For example, a simple lowpass filter for the gradients yields as an output ˆ ˆ − 1) − gˆ (n), S(n) = β S(n

S(0) = −ˆg(0)

(2.26)

where the choice of 0 ≤ β < 1 determines the bandwidth of the lowpass filter. When β is close to unity, the filter bandwidth is small and the effective averaging is performed over many gradient vectors. On the other hand, when β is small, the lowpass filter has a large bandwidth and, hence, it provides little averaging of the gradient vectors. With the filtered gradient vectors given by (2.26) in place of gˆ (n), we obtain the filtered version of the LMS algorithm given by 1 ˆ hM (n + 1) = hM (n) + S(n) 2

(2.27)

An analysis of the filtered-gradient LMS algorithm is given in Proakis (1974). Three other variations of the basic LMS algorithm given in (2.23) are obtained by using sign information contained in the error signal sequence e(n) and/or in the components of the signal vector XM (n). Hence, the three possible variations are ∗ hM (n + 1) = hM (n) + csgn[e(n)]XM (n)

(2.28)

∗ hM (n + 1) = hM (n) + e(n)csgn[XM (n)]

(2.29)

∗ hM (n + 1) = hM (n) + csgn[e(n)]csgn[XM (n)]

(2.30)

where csgn[x] is the complex sign function defined as  1 + j,   1 − j, csgn[x] = −1 + j,   −1 − j,

928

if Re(x) > 0 and Im(x) > 0 if Re(x) > 0 and Im(x) < 0 if Re(x) < 0 and Im(x) > 0 if Re(x) < 0 and Im(x) < 0

Adaptive Filters

and csgn[X] denotes the complex sign function applied to each element of the vector X. These three variations of the LMS algorithm may be called reduced complexity LMS algorithms, since multiplications are completely avoided in (2.30), and can be completely avoided in (2.28) and (2.29) by selecting  to be a power of 1/2. The price paid for the reduction in computational complexity is a slower convergence of the filter coefficients to their optimum values. Another version of the LMS algorithms, called a normalized LMS(NLMS) algorithm, that is frequently used in practice is given as hM (n + 1) = hM (n) +

 ∗ e(n)XM (n) ||XM (n)||2

(2.31)

By dividing the step size by the norm of the data vector XM (n), the NLMS algorithm is equivalent to employing a variable step size of the form (n) =

 ||XM (n)||2

(2.32)

Thus, the step size at each iteration is inversely proportional to the energy in the received data vector XM (n). This scaling is advantageous in adaptive filtering applications where the dynamic range of the input to the adaptive filter is large, as would be the case, for example, in the implementation of adaptive equalizers for slowly fading communication channels. In such applications, it may be advantageous to add a small positive constant to the denominator of (2.32) to avoid numerical instabilities that may result when the norm of XM (n) is small. Thus, another version of the NLMS algorithm may employ a variable step size of the form (n) =

 δ + ||XM (n)||2

(2.33)

where δ is a small positive number.

2.4

Properties of the LMS Algorithm

In this section, we consider the basic properties of the LMS algorithm given by (2.23). In particular, we focus on its convergence properties, its stability, and the excess noise generated as a result of using noisy gradient vectors in place of the actual gradient vectors. The use of noisy estimates of the gradient vectors implies that the filter coefficients will fluctuate randomly and, hence, an analysis of the characteristics of the algorithm should be performed in statistical terms. The convergence and stability of the LMS algorithm may be investigated by determining how the mean value of hM (n) converges to the optimum coefficients hopt . If we take the expected value of (2.23), we obtain ∗ (n)] h¯ M (n + 1) = h¯ M (n) + E[e(n)XM

= h¯ M (n) + [γ d − M h¯ M (n)] = (I − M )h¯ M (n) + γd

(2.34)

929

Adaptive Filters

where h¯ M (n) = E[hM (n)], and I is the identity matrix. The recursive relation in (2.34) may be represented as a closed-loop control system, as shown in Figure 2.1. The convergence rate and the stability of this closed-loop system are governed by our choice of the step-size parameter . To determine the convergence behavior, it is convenient to decouple the M simultaneous difference equations given in (2.34), by performing a linear transformation of the mean coefficient vector h¯ M (n). The appropriate transformation is obtained by noting that the autocorrelation matrix M is Hermitian and, hence, it can be represented (see Gantmacher (1960)) as M = UUH (2.35) where U is the normalized modal matrix of M and  is a diagonal matrix with diagonal elements λk , 0 ≤ k ≤ M − 1, equal to the eigenvalues of M . When (2.35) is substituted into (2.34), the latter may be expressed as 0 0 h¯ M (n + 1) = (I − )h¯ M (n) + γd0

(2.36)

0 where the transformed (orthogonalized) vectors are h¯ M (n) = UH h¯ M (n) and γd0 = H U γd . The set of M first-order difference equations in (2.36) are now decoupled. Their convergence and their stability are determined from the homogeneous equation 0 0 (2.37) h¯ M (n + 1) = (I − )h¯ M (n)

If we focus our attention on the solution of the kth equation in (2.37), we observe that h¯ 0 (k, n) = C(1 − λk )n u(n),

k = 0, 1, 2, . . . , M − 1

(2.38)

where C is an arbitrary constant and u(n) is the unit step sequence. Clearly, h¯ 0 (k, n) converges to zero exponentially, provided that |1 − λk | < 1 or, equivalently, 0 0 Em

fm (−1) = gm (−1) = km (−1) = 0 αm (−1) = 1,

(4.64)

α−1 (n) = α−1 (n − 1) = 1

Joint Process Estimation. The last step in the derivation is to obtain the least-squares estimate of the desired signal d(n) from the lattice. Suppose that the adaptive filter has m + 1 coefficients, which are determined to minimize the average weighted squared error n  Ᏹm+1 = w n−l |em+1 (l, n)|2 (4.65) l=0

where t (n)Xm+1 (l) em+1 (l, n) = d(l) − hm+1

(4.66)

ˆ n) = ht (n)Xm+1 (l) d(l, m+1

(4.67)

The linear estimate

which will be obtained from the lattice by using the residuals gm (n), is called the joint process estimate. From the results of Section 3.1, we have already established that the coefficients of the adaptive filter that minimize (4.65) are given by the equation hm+1 (n) = Pm+1 (n)Dm+1 (n)

(4.68)

We have also established that hm (n) satisfies the time-update equation given in (3.27). Now, let us obtain an order-update equation for hm (n). From (4.68) and (4.27), we have  hm+1 (n) =

Pm (n) 0 0 0



    1 Dm (n) bm (n)  H + b bm (n) 1 Dm+1 (n) ··· 1 Em (n)

(4.69)

We define a complex-valued scalar quantity δm (n) as   δm (n) = bH m (n) 1 Dm+1 (n)

(4.70)

Then, (4.69) may be expressed as  hm+1 (n) =

958

   δm (n) bm (n) hm (n) + b 0 1 Em (n)

(4.71)

Adaptive Filters

The scalar δm (n) satisfies a time-update equation that is obtained from the timeupdate equations for bm (n) and Dm (n), given by (4.51) and (3.17), respectively. Thus,   H ∗ ∗ δm (n) = bH m (n − 1) − Km (n)gm (n) 1 [wDm+1 (n − 1) + d(n)Xm+1 (n)]   ∗ = wδm (n − 1) + bH m (n − 1) 1 Xm+1 (n)d(n)  H   H  ∗ ∗ ∗ − wgm (n) Km (n) 0 Dm+1 (n − 1) − gm (n)d(n) Km (n) 0 Xm+1 (n)

(4.72)

But 

 ∗ ∗ H ∗ ∗ bH m (n − 1) 1 Xm+1 (n) = x (n − m) + bm (n − 1)Xm (n) = gm (n)

(4.73)

Also, 

H (n) Km

   t  Dm (n − 1) 1 0 Dm+1 (n − 1) = X (n)Pm (n − 1) 0 ··· w + µm (n) m 

(4.74)

1 = Xt (n)hm (n − 1) w + µm (n) m The last term in (4.72) may be expressed as 

H (n) 0 Km

  ∗  Xm 1 (n) ∗ = Xt (n)Pm (n − 1)Xm (n) ··· w + µm (n) m

(4.75)

µm (n) = w + µm (n) Upon substituting the results in (4.73–4.75) into (4.72), we obtain the desired timeupdate equation for δm (n) as ∗ (n)em (n) δm (n) = wδm (n − 1) + αm (n)gm

(4.76)

Order-update equations for αm (n) and gm (n) have already been derived. With e0 (n) = d(n), the order-update equation for em (n) is obtained as follows: t (n − 1)Xm (n) em (n) = em (n, n − 1) = d(n) − hm    Xm−1 (n)  t (n − 1) 0 = d(n) − hm−1 ···



 δm−1 (n − 1)  t bm−1 (n − 1) 1 Xm (n) b Em−1 (n − 1)

= em−1 (n) −

δm−1 (n − 1)gm−1 (n) b Em−1 (n − 1)

(4.77)

959

Adaptive Filters

dˆ(n) e1(n)







e2(n)



e3(n)









eM1(n)





eM(n)



0(n1)

1(n1)

2(n1)

M2(n1)

M1(n1)

E b0(n1)

E b1(n1)

E b2(n1)

EbM2(n1)

EbM1(n1)

gM1(n)

⯗ ⯗

d(n)

g0(n)

g1(n) Stage 1

gM2(n)

Stage 2

Stage M1

⯗ ⯗

x(n)

g2(n)

f0(n)

f1(n)

f2(n)

fM2(n)

fM1(n)

Figure 4.2 Adaptive RLS lattice-ladder filter.

Finally, the output estimate d(n) of the least-squares lattice is t ˆ d(n) = hm+1 (n − 1)Xm+1 (n)

(4.78)

t But hm+1 (n − 1) is not computed explicitly. By repeated substitution of the orderupdate equation for hm+1 (n) given by (4.71) into (4.78), we obtain the desired ˆ expression for d(n) in the form

ˆ d(n) =

M−1  k=0

δk (n − 1) gk (n) Ekb (n − 1)

(4.79)

ˆ In other words, the output estimate d(n) is a linear weighted sum of the backward residuals gk (n). The adaptive least-squares lattice/joint-process (ladder) estimator is illustrated in Figure 4.2. This lattice-ladder structure is mathematically equivalent to the RLS direct-form FIR filter. The recursive equations are sumarized in Table 5. This is called the a priori form of the RLS lattice-ladder algorithm in order to distinguish it from another form of the algorithm, called the a posteriori form, in which the coefficient vector hM (n) is used in place of hM (n−1) to compute the estimate d(n). In many adaptive filtering problems, such as channel equalization and echo cancellation, the a posteriori form cannot be used, because hM (n) cannot be computed prior to the computation of d(n). We now describe a number of modifications that can be made to the “conventional” RLS lattice-ladder algorithm given in Table 5.

960

Adaptive Filters

TABLE 5 A Priori Form of the RLS Lattice-Ladder Algorithm

Lattice predictor: Begin with n = 1 and compute the order updates for m = 0, 1, . . . , M − 2 ∗ km+1 (n − 1) = wkm+1 (n − 2) + αm (n − 2)fm (n − 1)gm (n − 2) f

᏷m+1 (n − 1) = − ᏷bm+1 (n − 1) = −

km+1 (n − 1) Emb (n − 2) ∗ (n − 1) km+1 f

Em (n − 1) f

fm+1 (n) = fm (n) + ᏷m+1 (n − 1)gm (n − 1) gm+1 (n) = gm (n − 1) + ᏷bm+1 (n − 1)fm (n) f

Em+1 (n − 1) = Emf (n − 1) − b (n − 1) = Emb (n − 2) − Em+1

αm+1 (n − 1) = αm (n − 1) −

|km+1 (n − 1)|2 Emb (n − 2) |km+1 (n − 1)|2 f

Em (n − 1) 2 αm (n − 1)|gm (n − 1)|2 Emb (n − 1)

Ladder filter: Begin with n = 1 and compute the order updates for m = 0, 1, . . . , M − 1 ∗ δm (n − 1) = wδm (n − 2) + αm (n − 1)gm (n − 1)em (n − 1)

ξm (n − 1) = −

δm (n − 1) Emb (n − 1)

em+1 (n) = em (n) + ξm (n − 1)gm (n)

Initialization α0 (n − 1) = 1, e0 (n) = d(n), f E0 (n)

=

E0b (n)

=

f wE0 (n

f0 (n) = g0 (n) = x(n)

− 1) + |x(n)|2

αm (−1) = 1, km (−1) = 0 Emb (−1) = Emf (0) = > 0; δm (−1) = 0

Modified RLS Lattice Algorithms. The recursive equations in the RLS lattice algorithm given in Table 5 are by no means unique. Modifications can be made to some of the equations without affecting the optimality of the algorithm. However, some modifications result in algorithms that are more numerically robust when fixedpoint arithmetic is used in the implementation of the algorithms. We give a number of basic relationships that are easily established from the above developments. First, we have a relationship between the a priori and a posteriori error residuals. A priori errors: t (n − 1)Xm (n − 1) fm (n, n − 1) ≡ fm (n) = x(n) + am

gm (n, n − 1) ≡ gm (n) = x(n − m) + btm (n − 1)Xm (n)

(4.80)

961

Adaptive Filters

A posteriori errors: t fm (n, n) = x(n) + am (n)Xm (n − 1)

gm (n, n) = x(n − m) + btm (n)Xm (n)

(4.81)

The basic relations between (4.80) and (4.81) are fm (n, n) = αm (n − 1)fm (n) gm (n, n) = αm (n)gm (n)

(4.82)

These relations follow easily by using (4.50) and (4.51) in (4.81). Second, we may obtain time-update equations for the least-squares forward and backward errors. For example, from (4.8) and (4.50) we obtain f t Em (n) = q(n) + am (n)Q∗m (n)  t  t = q(n) + am (n − 1) − Km (n − 1)fm (n) [wQ∗m (n − 1) + x ∗ (n)Xm (n − 1)] f = wEm (n − 1) + αm (n − 1)|fm (n)|2

(4.83)

Similarly, from (4.17) and (4.51) we obtain b b Em (n) = wEm (n − 1) + αm (n)|gm (n)|2

(4.84)

Usually, (4.83) and (4.84) are used in place of the sixth and seventh equations in Table 5. Third, we obtain a time-update equation for the Kalman gain vector, which is not explicitly used in the lattice algorithm, but which is used in the fast FIR filter algorithms. For this derivation, we also use the time-update equations for the forward and backward prediction coefficients given by (4.50) and (4.51). Thus, we have ∗ Km (n) = Pm (n)Xm (n)    x ∗ (n) 0 0 = ∗ 0 Pm−1 (n − 1) Xm−1 (n − 1)       1 1 x ∗ (n) H + f 1 am−1 (n) ∗ Xm−1 (n − 1) Em−1 (n) am−1 (n)     f ∗ (n, n) 0 1 = + m−1 f Km−1 (n − 1) Em−1 (n) am−1 (n)   Cm−1 (n) ≡ cmm (n)

962

(4.85)

Adaptive Filters

where, by definition, Cm−1 (n) consists of the first (m − 1) elements of Km (n) and cmm (n) is the last element. From (4.60), we also have the order-update equation for Km (n) as     g ∗ (n, n) bm−1 (n) Km−1 (n) + m−1 (4.86) Km (n) = 0 1 E b (n) m−1

By equating (4.85) to (4.86), we obtain the result cmm (n) =

∗ (n, n) gm−1

(4.87)

b Em−1 (n)

and, hence, Km−1 (n) + cmm (n)bm−1 (n) = Cm−1 (n)

(4.88)

By substituting for bm−1 (n) from (4.51) into (4.88), we obtain the time-update equation for the Kalman gain vector in (4.85) as Km−1 (n) =

Cm−1 (n) − cmm (n)bm−1 (n − 1) 1 − cmm (n)gm−1 (n)

(4.89)

There is also a time-update equation for the scalar αm (n). From (4.63), we have αm (n) = αm−1 (n) −

2 (n)|gm−1 (n)|2 αm−1 b Em−1 (n)

= αm−1 (n)[1 − cmm (n)gm−1 (n)]

(4.90)

A second relation is obtained by using (4.85) to eliminate K m−1 (n) in the expression for αm (n). Then, t αm (n) = 1 − Xm (n)Km (n) 

= αm−1 (n − 1) 1 −

∗ (n, n)fm−1 (n) fm−1 f

Em−1 (n)

 (4.91)

By equating (4.90) to (4.91), we obtain the desired time-update equation for αm (n) as   f ∗ (n,n)fm−1 (n) 1 − m−1 f   Em−1 (n)  (4.92) αm−1 (n) = αm−1 (n − 1)   1 − cmm (n)gm−1 (n) 

963

Adaptive Filters

Finally, we wish to distinguish between two different methods for updating the reflection coefficients in the lattice filter and the ladder part: the conventional (indirect) method and the direct method. In the conventional (indirect) method, f

᏷m+1 (n) = − ᏷bm+1 (n) = − ξm (n) = −

km+1 (n) b (n − 1) Em ∗ (n) km+1 f

Em (n) δm (n) b (n) Em

(4.93) (4.94) (4.95)

where km+1 (n) is time-updated from (4.58), δm (n) is updated according to (4.76), and f b Em (n) and Em (n) are updated according to (4.83) and (4.84). By substituting for km+1 (n) from (4.58) into (4.93), and using (4.84) and the eighth equation in Table 5, we obtain   b ∗ km+1 (n − 1) wEm (n − 2) (n − 1) αm (n − 1)fm (n)gm f − ᏷m+1 (n) = − b b b Em (n − 2) Em (n − 1) Em (n − 1)   αm (n − 1)|gm (n − 1)|2 f = ᏷m+1 (n − 1) 1 − b (n − 1) Em ∗ (n − 1) αm (n − 1)fm (n)gm b Em (n − 1)

− f

= ᏷m+1 (n − 1) −

∗ αm (n − 1)fm+1 (n)gm (n − 1) b Em (n − 1)

(4.96)

which is a formula for directly updating the reflection coefficients in the lattice. Similarly, by substituting (4.58) into (4.94), and using (4.83) and the eighth equation in Table 5, we obtain ᏷bm+1 (n) = ᏷bm+1 (n − 1) −

αm (n − 1)fm∗ (n)gm+1 (n) f

Em (n)

(4.97)

Finally, the ladder gain can also be updated directly according to the relation ξm (n) = ξm (n − 1) −

∗ αm (n)gm (n)em+1 (n) b Em (n)

(4.98)

The RLS lattice-ladder algorithm that uses the direct update relations in (4.96– 4.98) and (4.83–4.84) is listed in Table 6. An important characteristic of the algorithm in Table 6 is that the forward and backward residuals are fed back to time-update the reflection coefficients in the lattice stage, and em+1 (n) is fed back to update the ladder gain ξm (n). For this reason, this RLS lattice-ladder algorithm has been called the error-feedback form. A similar form can be obtained for the a posteriori RLS lattice-ladder algorithm. For more details on the error-feedback form of RLS lattice-ladder algorithms, the interested reader is referred to Ling, Manolakis, and Proakis (1986).

964

Adaptive Filters

TABLE 6

Direct Update (Error-Feedback) Form of the A Priori RLS Lattice-Ladder

Algorithm

Lattice predictor: Begin with n = 1 and compute the order updates for m = 0, 1, . . . , M − 2 ∗ αm (n − 2)fm+1 (n − 1)gm (n − 2) f f ᏷m+1 (n − 1) = ᏷m+1 (n − 2) − b Em (n − 2) ᏷bm+1 (n − 1) = ᏷bm+1 (n − 2) − fm+1 (n) = fm (n) +

αm (n − 2)fm∗ (n − 1)gm+1 (n − 1)

f ᏷m+1 (n

f

Em (n − 1) − 1)gm (n − 1)

gm+1 (n) = gm (n − 1) + ᏷bm+1 (n − 1)fm (n) f

f

Em+1 (n − 1) = wEm+1 (n − 2) + αm+1 (n − 2)|fm+1 (n − 1)|2 αm+1 (n − 1) = αm (n − 1) −

2 (n − 1)|gm (n − 1)|2 αm Emb (n − 1)

b b (n − 1) = wEm+1 (n − 2) + αm+1 (n − 1)|gm+1 (n − 1)|2 Em+1

Ladder filter: Begin with n = 1 and compute the order updates m = 0, 1, . . . , M − 1 ∗ αm (n − 1)gm (n − 1)em+1 (n − 1) ξm (n − 1) = ξm (n − 2) − Emb (n − 1) em+1 (n) = em (n) + ξm (n − 1)gm (n)

Initialization α0 (n − 1) = 1, e0 (n) = d(n), f E0 (n)

=

E0b (n)

αm (−1) = 1, Emb (−1)

=

=

f wE0 (n

f0 (n) = g0 (n) = x(n)

− 1) + |x(n)|2

᏷fm (−1) = ᏷bm (−1) = 0

Emf (0)

= >0

Fast RLS Algorithms. The two versions of the fast RLS algorithms given in Section 3.3 follow directly from the relationships that we have obtained in this section. In particular, we fix the size of the lattice and the associated forward and backward predictors at M − 1 stages. Thus, we obtain the first seven recursive equations in the two versions of the algorithm. The remaining problem is to determine the time-update equation for the Kalman gain vector, which was determined in (4.85– 4.89). In version B of the algorithm, given in Table 3, we used the scalar αm (n) to reduce the computations from 10M to 9M . Version A of the algorithm, given in Table 2, avoids the use of this parameter. Since these algorithms provide a direct updating of the Kalman gain vector, they have been called fast Kalman algorithms (for reference, see Falconer and Ljung (1978) and Proakis (1989)). Further reduction of computational complexity to 7M is possible by directly updating the following alternative (Kalman) gain vector (see Carayannis, Manolakis, and Kalouptsidis (1983)) defined as ∗ ˜ M (n) = 1 PM (n − 1)XM (n) K w

(4.99)

965

Adaptive Filters

Several fast algorithms using this gain vector have been proposed, with complexities ranging from 7M to 10M . Table 7 lists the FAEST (Fast A Posteriori Error Sequential Technique) algorithm with a computational complexity 7M (for a derivation, see Carayannis, Manolakis, and Kalouptsidis (1983; 1986) and Problem 7). In general, the 7M fast RLS algorithms and some variations are very sensitive to round-off noise and exhibit instability problems (Falconer and Ljung (1978), Carayannis, Manolakis, and Kalouptsidis (1983; 1986), and Cioffi and Kailath (1984)). The instability problem in the 7M algorithms has been addressed by Slock and Kailath (1988; 1991), and modifications have been proposed that stabilize these algorithms. The resulting stabilized algorithms have a computational complexity ranging from 8M to 9M . Thus, their computational complexity is increased by a relatively small amount compared to the unstable 7M algorithms. To understand the stabilized fast RLS algorithms, we begin by comparing the fast RLS algorithm given in Table 3 and the FAEST algorithm in Table 7. As TABLE 7 FAEST Algorithm t fM−1 (n) = x(n) + aM−1 (n − 1)XM−1 (n − 1)

f¯M−1 (n, n) =

fM−1 (n) α¯ M−1 (n − 1)

¯ M−1 (n − 1)f¯M−1 (n, n) aM−1 (n) = aM−1 (n − 1) − K ∗ EM−1 (n) = wEM−1 (n − 1) + fM−1 (n)fM−1 (n, n)  ¯      ∗ fM−1 (n) 0 1 ¯ M (n) ≡ CM−1 (n) = K + f ¯ M−1 (n − 1) c¯ MM (n) K wE (n − 1) aM−1 (n − 1) f

f

M−1

gM−1 (n) =

b −wEM−1 (n



∗ 1)¯cMM (n)

¯ M−1 (n) − bM−1 (n − 1)¯cMM (n) ¯ M−1 (n) = C K α¯ M (n) = α¯ M−1 (n − 1) +

|fM−1 (n)|2 f

wEM−1 (n − 1)

α¯ M−1 (n) = α¯ M (n) + gM−1 (n)¯cMM (n) gM−1 (n) g¯ M−1 (n, n) = α¯ M−1 (n) b ∗ b (n) = wEM−1 (n − 1) + gM−1 (n)g¯ M−1 (n, n) EM−1

¯ M−1 (n)g¯ M−1 (n, n) bM−1 (n) = bM−1 (n − 1) + K t eM (n) = d(n) − hM (n − 1)XM (n) e¯M (n, n) =

eM (n) α¯ M (n)

¯ M (n)e¯M (n, n) hM (n) = hM (n − 1) + K

Initialization: Set all vectors to zero f

b EM−1 (−1) = EM−1 (−1) = > 0

α¯ M−1 (−1) = 1

966

Adaptive Filters

indicated, there are two major differences between these two algorithms. First, the FAEST algorithm uses the alternative (Kalman) gain vector instead of the Kalman gain vector. Second, the fast RLS algorithm computes the a priori backward prediction error gM−1 (n) through FIR filtering using the backward prediction coefficient vector bm−1 (n − 1), whereas the FAEST algorithm computes the same quantity through a scalar operation by noticing that the last element of the alternative gain b vector, c˜ MM (n), is equal to −wEM−1 gM−1 (n). Since these two algorithms are algebraically equivalent, the backward prediction errors calculated in different ways should be identical if infinite precision is used in the computation. Practically, when finite-precision arithmetic is used, the backward prediction errors computed using different formulas are only approximately equal. In what follows, we denote them (f ) (s) by gM−1 and gM−1 (n), respectively. The superscripts (f ) and (s) indicate that they are computed using the filtering approach and scalar operation, respectively. There are other quantities in the algorithms that can also be computed in different ways. In particular, the parameter αM−1 (n) can be computed from the vector ˜ M−1 (n) and XM−1 (n) as quantities K ˜ t (n)XM−1 (n) αM−1 (n) = 1 + K M−1

(4.100)

(f )

(s) (n), respecor from scalar quantities. We denote these values as α˜ M−1 (n) and α˜ M−1 (f ) ˜ tively. Finally, the last element of KM (n), denoted as c˜ MM (n), may be computed from the relation (f ) −gM−1 (n) (f ) (4.101) c˜ MM (n) = b wEM−1 (n − 1) (f )

(f )

(s) The two quantities in each of the three pairs [gM−1 (n), gM−1 (n)], [αM−1 (n), (f )

(s) (s) αM−1 (n)], and [˜cMM (n), c˜ MM (n)] are algebraically equivalent. Hence, either of the two quantities or their linear combination (of the form kβ (s) + (1 − k)β (f ) , where β represents any of the three parameters) are algebraically equivalent to the original quantities, and may be used in the algorithm. Slock and Kailath (1988; 1991) found that by using the appropriate quantity or its linear combination in the fast RLS algorithm, it was sufficient to correct for the positive feedback inherent in the fast RLS algorithms. Implementation of this basic notion leads to the stabilized fast RLS algorithm given in Table 8. We observe from Table 8 that the stabilized fast RLS algorithm employs constants ki , i = 1, 2, . . . , 5, to form five linear combinations of the three pairs of quantities just described. The best values of the ki found by Slock and Kailath resulted from computer search, and are given as k1 = 1.5, k2 = 2.5, k3 = 1, k4 = 0, k5 = 1. When ki = 0 or 1, we use only one of the quantities in the linear combination. Hence, some of the parameters in the three pairs need not be computed. It was also found that (f ) the stability of the algorithm is only slightly affected if αM−1 (n) is not used. These simplifications result in the algorithm given in Table 9, which has a complexity of 8M and is numerically stable. The performance of the stabilized fast RLS algorithms depends highly on proper (f ) initialization. On the other hand, an algorithm that uses gM−1 (n) in its computations

967

Adaptive Filters

TABLE 8 The Stabilized Fast RLS Algorithm t fM−1 (n) = x(n) + aM−1 (n − 1)XM−1 (n − 1)

fM−1 (n, n) =

fM−1 (n) α¯ M−1 (n − 1)

¯ M−1 (n − 1)fM−1 (n, n) aM−1 (n) = aM−1 (n − 1) − K ∗ (n) fM−1

c¯ M1 (n) =

f

wEM−1 (n − 1)       ¯ CM−1 (n) 1 0 = ¯ + c¯ M1 (n) (s) c¯ MM (n) KM−1 (n − 1) aM−1 (n − 1) (f )

gM−1 (n) = x(n − M + 1) + btM−1 (n − 1)XM−1 (n) (f )∗

gM−1 (n)

(f )

c¯ MM (n) = −

b wEM−1 (n − 1) (f )

(s) c¯ MM (n) = k4 c¯ MM (n) + (1 − k4 )¯cMM (n)  ¯  ¯ M (n) = CM−1 (n) K c¯ MM (n) (s) (s)∗ b gM−1 (n) = −wEM−1 (n − 1)¯cMM (n) (f )

(i) (s) gM−1 (n) = ki gM−1 (n) + (1 − ki )gM−1 (n),

i = 1, 2, 5

¯ M−1 (n) − bM−1 (n − 1)¯cMM (n) ¯ M−1 (n) = C K α¯ M (n) = α¯ M−1 (n − 1) + c¯ M1 (n)fM−1 (n) (s) α¯ M−1 (n)

(s) (s) = α¯ M (n) + gM−1 (n)¯cMM (n)

(f ) ¯ t (n)XM−1 (n) α¯ M−1 (n) = 1 + K M−1 (f )

(s) α¯ M−1 (n) = k3 α¯ M−1 (n) + (1 − k3 )α¯ M−1 (n) ∗ EM−1 (n) = wEM−1 (n − 1) + fM−1 (n)fM−1 (n, n)  1 1 |¯cM1 (n)|2 1 = − (s) f f w EM−1 (n − 1) EM−1 (n) α¯ M−1 (n) f

 or,

f

(i) gM−1 (n, n) =

(i) (n) gM−1 , α¯ M−1 (n)

i = 1, 2

¯ M−1 (n)g (1) (n, n) bM−1 (n) = bM−1 (n − 1) + K M−1 (2) (2)∗ b b (n) = wEM−1 (n − 1) + gM−1 (n)gM−1 (n, n) EM−1 t (n − 1)XM (n) eM (n) = d(n) − hM

eM (n, n) =

eM (n) α¯ M (n)

¯ M (n)eM (n, n) hM (n) = hM (n − 1) + K

968

Adaptive Filters

TABLE 9 A Simplified Stabilized Fast RLS Algorithm t fM−1 (n) = x(n) + aM−1 (n − 1)XM−1 (n − 1)

fM−1 (n, n) =

fM−1 (n) α¯ M−1 (n − 1)

¯ M−1 (n − 1)fM−1 (n, n) aM−1 (n) = aM−1 (n − 1) − K c¯ M1 (n) = ¯ M (n) ≡ K

 ¯  CM−1 (n) c¯ MM (n)

∗ (n) fM−1 f

wEM−1 (n − 1)     ∗ (n) fM−1 0 1 = ¯ + f KM−1 (n − 1) wEM−1 (n − 1) aM−1 (n − 1)

(f )

gM−1 (n) = x(n − M + 1) + btM−1 (n − 1)XM−1 (n) (s) b ∗ gM−1 (n) = −wEM−1 (n − 1)¯cMM (n) (f )

(i) (s) (n) = ki gM−1 (n) + (1 − ki )gM−1 (n), gM−1

i = 1, 2

¯ M−1 (n) − bM−1 (n − 1)¯cMM (n) ¯ M−1 (n) = C K α¯ M (n) = α¯ M−1 (n − 1) + c¯ M1 (n)fM−1 (n) (f )

α¯ M−1 (n) = α¯ M (n) + gM−1 (n)¯cMM (n) ∗ EM−1 (n) = wEM−1 (n − 1) + fM−1 (n)fM−1 (n, n) f

f

(i) gM−1 (n, n) =

(i) (n) gM−1 , α¯ M−1 (n)

i = 1, 2

¯ M−1 (n)g (1) (n, n) bM−1 (n) = bM−1 (n − 1) + K M−1 (2) (2)∗ b b EM−1 (n) = wEM−1 (n − 1) + gM−1 (n)gM−1 (n, n) t (n − 1)XM (n) eM (n) = d(n) − hM

eM (n, n) =

eM (n) α¯ M (n)

¯ M (n)eM (n, n) hM (n) = hM (n − 1) + K

is not critically affected by proper initialization (although, enventually, it will di(f ) (s) verge). Consequently, we may initially use gM−1 (n) in place of gM−1 (n) (or their linear combination) for the first few hundred iterations, and then switch to the form for the stabilized fast RLS algorithm. By doing so, we obtain a stabilized fast RLS algorithm that is also insensitive to initial conditions.

4.2

Other Lattice Algorithms

Another type of RLS lattice algorithm is obtained by normalizing the forward  and  f

b

backward prediction errors through division of the errors by Em (n) and Em (n), √ √ respectively, and multiplication by αm (n − 1) and αm (n), respectively. The resulting lattice algorithm is called a square-root or angle-and-power normalized RLS

969

Adaptive Filters

lattice algorithm. This algorithm has a more compact form than the other forms of RLS lattice algorithms. However, the algorithm requires many square-root operations, which can be computationally complex. This problem can be solved by using CORDIC processors, which compute a square root in N clock cycles, where N is the number of bits of the computer word length. A description of the squareroot/normalized RLS lattice algorithm and the CORDIC algorithm is given in the book by Proakis et al. (2002). It is also possible to simplify the computational complexity of the RLS algorithms described in the previous section at the expense of compromising the convergence rate. One such algorithm is called the gradient-lattice algorithm. In this algorithm each stage of the lattice filter is characterized by the output–input relations fm (n) = fm−1 (n) − km (n)gm−1 (n − 1) ∗ gm (n) = gm−1 (n − 1) − km (n)fm−1 (n)

(4.102)

where km (n) is the reflection coefficient in the mth stage of the lattice and fm (n) and gm (n) are the forward and backward residuals. This form of the lattice filter is identical to the Levinson-Durbin algorithm, except that now km (n) is allowed to vary with time so that the lattice filter adapts to the time variations in the signal statistics. The reflection coefficients {km (n)} may be optimized by employing the method of least squares, which results in the solution n

∗ n−l fm−1 (l)gm−1 (l − 1) l=0 w , n−l 2 [|fm−1 (l)| + |gm−1 (l − 1)|2 ] l=0 w

km (n) = n

2

m = 1, 2, . . . M − 1

(4.103)

These coefficients can also be updated recursively in time. The ladder coefficients are computed recursively in time by employing an LMS-type algorithm that is obtained by applying the mean-square-error criterion. A description of this algorithm is given in the paper by Griffiths (1978) and in the book by Proakis et al. (2002).

4.3

Properties of Lattice-Ladder Algorithms

The lattice algorithms that we have derived in the two previous subsections have a number of desirable properties. In this subsection, we consider the properties of these algorithms and compare them with the corresponding properties of the LMS algorithm and the RLS direct-form FIR filtering algorithms. Convergence Rate. The RLS lattice-ladder algorithms basically have the same convergence rate as the RLS direct-form FIR filter structures. This characteristic behavior is not surprising, since both filter structures are optimum in the least-squares sense. Although the gradient lattice algorithm retains some of the optimal characteristics of the RLS lattice, nevertheless the former is not optimum in the least-squares sense and, hence, its convergence rate is slower.

970

Adaptive Filters

Log of output mean-square error

Channel-correlation matrix Eigenvalue ratio  11 11-tap equalizer, noise variance  0.001 0.0 Gradient lattice algorithm 1.0

Least-squares lattice algorithm

Gradient algorithm (LMS)

2.0

3.0

Optimum

0

100

200

300

400 500 600 Number of iterations

700

800

900

Figure 4.3 Learning curves for RLS lattice, gradient lattice, and LMS algorithms for adaptive equalizer of length M = 11. (From Digital Communications by John G. Proakis. ©1989 by McGrawHill Book Company. Reprinted with permission of the publisher.)

For comparison purposes, Figures 4.3 and 4.4 illustrate the learning curves for an adaptive equalizer of length M = 11, implemented as an RLS lattice-ladder filter, a gradient lattice-ladder filter, and a direct-form FIR filter using the LMS algorithm, for a channel autocorrelation matrix that has eigenvalue ratios of λmax /λmin = 11 and λmax /λmin = 21, respectively. From these learning curves, we observe that the gradient lattice algorithm takes about twice as many iterations to converge as the optimum RLS lattice algorithm. Furthermore, the gradient lattice algorithm provides significantly faster convergence than the LMS algorithm. For both lattice structures, the convergence rate does not depend on the eigenvalue spread of the correlation matrix. Computational Requirements. The RLS lattice algorithms described in the previous subsection have a computational complexity that is proportional to M . In contrast, the computational complexity of the RLS square-root algorithms is proportional to M 2 . On the other hand, the direct-form fast algorithms, which are a derivative of the lattice algorithm, have a complexity proportional to M , and they are a little more efficient than the lattice-ladder algorithms. In Figure 4.5 we illustrate the computational complexity (number of complex multiplications and divisions) of the various adaptive filtering algorithms that we have described. Clearly, the LMS algorithm requires the fewest computations. The fast RLS algorithms in Tables 3 and 9 are the most efficient of the RLS algorithms shown, closely followed by the gradient lattice algorithm, then the RLS lattice algorithms and, finally, the square-root algorithms. Note that for small values of M , there is little difference in complexity among the rapidly convergent algorithms. Numerical Properties. In addition to providing fast convergence, the RLS and gradient lattice algorithms are numerically robust. First, these lattice algorithms are

971

Adaptive Filters

Log of output mean-square error

Channel-correlation matrix Eigenvalue ratio  21 11-tap equalizer, noise variance  0.001 0.0 Gradient lattice algorithm Least-squares lattice algorithm

1.0

Gradient algorithm (LMS) Optimum

2.0

3.0

0

100

200

300

400 500 600 Number of iterations

700

800

900

Figure 4.4 Learning curves for RLS lattice, gradient lattice, and LMS algorithms for adaptive equalizer of length M = 11. (From Digital Communications by John G. Proakis. ©1989 by McGraw-Hill Book Company. Reprinted with permission of the publisher.)

Number of complex multiplications and divisions

700 Direct-form RLS square-root algorithm

600

500 RLS lattice-ladder algorithm

400

Gradient lattice-ladder

300

Fast RLS algorithm

200

100 LMS algorithm 0

5

10

15

20

25

M ~ Length of filter

Figure 4.5 Computational complexity of adaptive filter algo-

rithms.

972

M

Adaptive Filters

TABLE 10 Numerical Accuracy, in Terms of Output MSE for Channel with λmax /λmin = 11

and w = 0.975, MSE × 10−3

Algorithm Number of bits

RLS

Fast

Conventional

Error feedback

(including sign)

square root

RLS

RLS lattice

RLS lattice

16

2.17

2.17

2.16

2.16

2.30

13 11

2.33 6.14

2.21 3.34

3.09 25.2 365

2.22 3.09 31.6

2.30 19.0 311

9 a Algorithm

a

75.3

LMS

did not converge.

numerically stable, which means that the output estimation error from the computational procedure is bounded when a bounded error signal is introduced at the input. Second, the numerical accuracy of the optimum solution is also relatively good when compared to the LMS and the RLS direct-form FIR algorithms. For purposes of comparison, we illustrate in Table 10 the steady-state average squared error or (estimated) minimum MSE obtained through computer simulation from the two RLS lattice algorithms and the direct-form FIR filter algorithms described in Section 2. The striking result in Table 10 is the superior performance obtained with the RLS lattice-ladder algorithm, in which the reflection coefficients and the ladder gain are updated directly according to (4.96–4.98). This is the error-feedback form of the RLS lattice algorithm. It is clear that the direct updating of these coefficients is significantly more robust to round-off errors than all the other adaptive algorithms, including the LMS algorithm. It is also apparent that the two-step process used in the conventional RLS lattice algorithm to estimate the reflection coefficients is not as accurate. Furthermore, the estimation errors that are generated in the coefficients at each stage propagate from stage to stage, causing additional errors. The effect of changing the weighting factor w is illustrated in the numerical results given in Table 11. In this table, we give the minimum (estimated) MSE TABLE 11 Numerical Accuracy, in Terms of Output MSE, of A Priori Least-Squares Lattice Algorithm with Different Values of the Weighting Factor w, MSE, × 10−3

Algorithm w = 0.99 Number of bits with sign 16 13 11 9

w = 0.975

Conventional

Error feedback

2.14 7.08 39.8 750

2.08 2.11 3.88 44.1

w = 0.95

Conventional

Error feedback

Conventional

Error feedback

2.18 3.09 25.2 365

2.16 2.22 3.09 31.6

2.66 3.65 15.7 120

2.62 2.66 2.78 15.2

973

Adaptive Filters

obtained with the conventional and error-feedback forms of the RLS lattice algorithm. We observe that the output MSE decreases with an increase in the weighting factor when the precision is high (13 bits and 16 bits). This reflects the improvement in performance obtained by increasing the observation interval. As the number of bits of precision is decreased, we observe that the weighting factor should also be decreased in order to maintain good performance. In effect, with low precision, the effect of a longer averaging time results in a larger round-off noise. Of course, these results were obtained with time-invariant signal statistics. If the signal statistics are time-variant, the rate of the time variations will also influence the choice of w. In the gradient lattice algorithm, the reflection coefficients and the ladder gains are also updated directly. Consequently, the numerical accuracy of the gradient lattice algorithm is comparable to that obtained with the direct update form of the RLS lattice. Analytical and simulation results on numerical stability and numerical accuracy in fixed-point implementation of these algorithms can be found in Ling and Proakis (1984), Ling, Manolakis, and Proakis (1985; 1986), Ljung and Ljung (1985), and Gardner (1984). Implementation Considerations. As we have observed, the lattice filter structure is highly modular and allows for the computations to be pipelined. Because of the high degree of modularity, the RLS and gradient lattice algorithms are particularly suitable for implementation in VLSI. As a result of this advantage in implementation and the desirable properties of stability, excellent numerical accuracy, and fast convergence, we anticipate that adaptive filters will be increasingly implemented as lattice-ladder structures in the near future.

5

Summary and References We have presented adaptive algorithms for direct-form FIR and lattice filter structures. The algorithms for the direct-form FIR filter consisted of the simple LMS algorithm due to Widrow and Hoff (1960) and the direct-form, time-recursive leastsquares algorithms, including the conventional RLS form given by (3.23–3.27), the square-root RLS forms described by Bierman (1977), Carlson and Culmone (1979), and Hsu (1982), and the RLS fast Kalman algorithms, one form of which was described by Falconer and Ljung (1978), and other forms later derived by Carayannis, Manolakis, and Kalouptsidis (1983), Proakis (1989), and Cioffi and Kailath (1984). Of these algorithms, the LMS algorithm is the simplest. It is used in many applications where its slow convergence is adequate. Of the direct-form RLS algorithms, the square-root algorithms have been used in applications where fast convergence is required. The algorithms have good numerical properties. The family of stabilized fast RLS algorithms is very attractive from the viewpoint of computational efficiency. Methods to avoid instability due to round-off errors have been proposed by Hsu (1982), Cioffi and Kailath (1984), Lin (1984), Eleftheriou and Falconer (1987), and Slock and Kailath (1988; 1991). The adaptive lattice-ladder filter algorithm derived in this chapter is the optimum RLS lattice-ladder algorithms (in both conventional and error-feedback form). Only the a priori form of the lattice-ladder algorithm was derived, which is the form most

974

Adaptive Filters

often used in applications. In addition, there is an a posteriori form of the RLS lattice-ladder algorithms (both conventional and error-feedback), as described by Ling, Manolakis, and Proakis (1986). The error-feedback form of the RLS latticeladder algorithm has excellent numerical properties, and is particularly suitable for implementation in fixed-point arithmetic and in VLSI. In the direct-form and lattice RLS algorithms, we used exponential weighting into the past in order to reduce the effective memory in the adaptation process. As an alternative to exponential weighting, we may employ finite-length uniform weighting into the past. This approach leads to the class of finite-memory RLS direct-form and lattice structures described in Cioffi and Kalaith (1985) and Manolakis, Ling, and Proakis (1987). In addition to the various algorithms that we have presented in this chapter, there has been considerable research into efficient implementation of these algorithms using systolic arrays and other parallel architectures. The reader is referred to Kung (1982), and Kung, Whitehouse, and Kailath (1985).

Problems 1

2

3

4 5 6 7

Use the least-squares criterion to determine the equations for the parameters of the FIR filter model in Figure 1.2, when the plant output is corrupted by additive noise, w(n). Determine the equations for the coefficients of an adaptive echo canceler based on the least-squares criterion. Use the configuration in Figure 1.8 and assume the presence of a near-end echo only. If the sequences w1 (n), w2 (n), and w3 (n) in the adaptive noise-canceling system shown in Figure 1.14 are mutually uncorrelated, determine the expected value of the estimated correlation sequences rvv (k) and ryv (k) contained in (1.26). Prove the result in (4.34). Derive the equation for the direct update of the ladder gain given by (4.98). Derive the equation for the reflection coefficients in a gradient lattice-algorithm given in (4.103). Derive the FAEST algorithm given in Table 7 by using the alternative Kalman gain vector ∗ ˜ M (n) = 1 PM (n − 1)XM K (n) w

instead of the Kalman gain vector KM (n). 8 The tap-leaky LMS algorithm proposed by Gitlin, Meadors, and Weinstein (1982) may be expressed as ∗ hM (n + 1) = whM (n) + e(n)XM (n)

where 0 < w < 1,  is the step size, and XM (n) is the data vector at time n. Determine the condition for the convergence of the mean value of hM (n).

975

Adaptive Filters

9

The tap-leaky LMS algorithm in Problem 8 may be obtained by minimizing the cost function E (n) = |e(n)|2 + c||hM (n)||2 where c is a constant and e(n) is the error between the desired filter output and the actual filter output. Show that the minimization of E (n) with respect to the filter coefficient vector hM (n) leads to the following tap-leaky LMS algorithm. ∗ hM (n + 1) = (1 − c)hM (n) + e(n)XM (n)

10 11

12

For the normalized LMS algorithm given by (2.31), determine the range of values of step size  to ensure the stability of the algorithm in the mean-square-error sense. By using the alternative Kalman gain vector given in Problem 8, modify the a priori fast least-squares algorithms given in Tables 2 and 3, and thus reduce the number of computations. Consider the random process x(n) = gv(n) + w(n),

n = 0, 1, . . . , M − 1

where v(n) is a known sequence, g is a random variable with E[g] = 0, and E[g 2 ] = G. The process w(n) is a white noise sequence with γww (m) = σw2 δ(m) Determine the coefficients of the linear estimator for g , that is, gˆ =

M−1 

h(n)x(n)

n=0

that minimizes the mean-square error E = e[(g − g) ˆ 2] 13 Recall that an FIR filter can be realized in the frequency-sampling form with system function M−1 Hk 1 − z−M  H (z) = j M 1 − e 2πk/M z−1 k=0

= H1 (z)H2 (z) where H1 (z) is the comb filter and H2 (z) is the parallel bank of resonators. (a) Suppose that this structure is implemented as an adaptive filter using the LMS algorithm to adjust the filter (DFT) parameters Hk . Give the time-update equation for these parameters. Sketch the adaptive filter structure. (b) Suppose that this structure is used as an adaptive channel equalizer, in which the desired signal is d(n) =

M−1  k=0

976

Ak cos ωk n,

ωk =

2π k M

Adaptive Filters

14

With this form for the desired signal, what advantages are there in the LMS adaptive algorithm for the DFT coefficients Hk over the direct-form structure with coefficients h(n)? (Hint: Refer to Proakis (1970).) Consider the performance index J = h2 − 40h + 28 Suppose that we search for the minimum of J by using the steepest-descent algorithm 1 h(n + 1) = h(n) − g(n) 2 where g(n) is the gradient. (a) Determine the range of values of  that provides an overdamped system for the adjustment process. (b) Plot the expression for J as a function of n for a value of  in this range.

15 Consider the noise-canceling adaptive filter shown in Figure 1.14. Assume that the additive noise processes are white and mutually uncorrelated, with equal variances σw2 . Suppose that the linear system has a known system function H (z) =

16

1 1 − 21 z−1

Determine the optimum weights of a three-tap noise canceler that minimizes the MSE. Determine the coefficients a1 and a2 for the linear predictor shown in Figure P16, given that the autocorrelation γxx (n) of the input signal is γxx (m) = a |m| ,

0